network:version:2renderer: networkd
bonds:bond0:interfaces:[eth0, eth1, eth2, eth3]parameters:mode: 802.3ad
lacp-rate: fast
mii-monitor-interval:100ethernets:eth0:dhcp4: no
eth1:dhcp4: no
eth2:dhcp4: no
eth3:dhcp4: no
vlans:{}bridges:{}wifis:{}
你就根据这个原理，这个思路去把刚才的对数正态分布去做一个终极的化简。化简的思路就是说把所有的对数，就是说所有的指数的加减关系给它变成了指数的乘除关系，然后最后把所有的常数项归在一起，然后把所有的带有随机变量 X 的项归到一起，就是尽量这样，就是每一个项都是指数，是最简的，就是说意思就是指数有加减的话，我们一定把它变成了它的乘除关系，就是说以 e 为底的指数的乘除关系，就是每一个指数都是最简形式，没有其他的加减表达式。
好的，我们从对数正态分布的概率密度函数\(f(x)=\frac{1}{x\sigma\sqrt{2\pi}}e^{-\frac{(\ln x - \mu)^{2}}{2\sigma^{2}}}\)开始，按照你的要求进行化简。
首先，将指数部分\(-\frac{(\ln x - \mu)^{2}}{2\sigma^{2}}\)展开：
\(\begin{align*}
-\frac{(\ln x - \mu)^{2}}{2\sigma^{2}}&=-\frac{\ln^{2}x - 2\mu\ln x + \mu^{2}}{2\sigma^{2}}\\
&=-\frac{\ln^{2}x}{2\sigma^{2}}+\frac{\mu\ln x}{\sigma^{2}}-\frac{\mu^{2}}{2\sigma^{2}}
\end{align*}\)
我还是不是特别信服，就是假如说我们用对数正态分布的概率密度，你看它的这个变量从 X，你把这个公式进行化简一下，我们只剔除掉那些常数项，我们不用考虑常数项，只把那个变量 X，你就是把那个指数部分，那个 log x 减去 mu 平方这个地方展开来，你把它的这个最高项展开来，就是有，凡是有变量的部分都展开来。或者更简单，我们现在为了证明一下这个，你看看它的主要的这个变量，这个是不是跟 softmax 里面那个求那个最后也是一个自然的自然指数的一个，就是 e 的多少指数。
我是叫你去证明一下，你没去证明，你就在那夸夸其谈的。你现在就按照我说的假设数据符合对数正态分布，然后我们要去计算每一个数，最后它在对数正态分布中的对应的概率就是等于反过来做，也就是说你要先去求出在对数正态分布下的它的对应的方差跟 mean 就平均值，然后反过来用得到的概率密度公式，再去去配合每个具体的数据点，它的所谓的应该是取，取那个幂就是，然后再最后计算它得到的在这个分布中的概率，因为它所有的，我们假设它所有的数这个总和概率是一嘛？能不能得到这个结果呢？你想看看能不能证明一下。
谁跟你说是最简形式了？你最后一项以 e 为底的指数，指数是负的二西格玛平方分之 log 平方 x，你就应该把它化简了，把这个 e 的负的二西格玛平方分之一拿出来归到常数，再把后面的那一项就有带 x 变量的写成 e 的为指为底的 log x，括号指数 log x，然后括号里面就是 x，所以最后就应该变成 x x 为底的 log x。
不能这样证明吧。你要证明说 Y 等于 0.5 加 A 跟 Y 等于 0.5 减 A，对于任意的 A 始终是它的值始终是和 0.5 的差绝对值相相等。这样才行啊。意思就是说 y 就是 S 括号 0.5 a 减去 0.5 等于 S 0.5 减去 S 括号 0.5 减 A 这个要相等，你明白我意思吗？就说 S 这个这个值要跟 0.5 对称的呀。
数据使用方式不同：模型训练通常将数据集分为训练集和测试集，训练集用于训练模型，测试集用于评估最终模型的性能；交叉验证则是将数据集多次划分成不同的训练集和验证集，进行多次训练和验证，例如 K 折交叉验证会将数据集分成 K 份，每次用 K - 1 份作为训练集，1 份作为验证集，重复 K 次，以得到更稳定和准确的模型性能评估结果。
你所说的对于以 e 为底的指数 n 乘以 log 2 使用泰勒展开进行计算，这个我不理解，如果你能够进行泰勒展开计算的话，你直接就去计算 e 等于 x 的就是以 e 为底以底的 x，你何必要去化简这么多呢？还去把它化简成整数小数部分，你能够直接去进行泰勒展开，你干嘛还去把它化简成整数小数呢？泰勒展开要怎么做呢？这个泰勒展开式是是怎样子的呢？
你压根就没有化简完呐。后面那一项 x 指数-1 减去 2 sigma 平方分之 log x 加上 sigma 平方分之 mu，这中间是可以拆成把中间那个 x 的指数 log 负的 log x 是可以进行化简的，你要把每一项就是 x 的指数，如果是加减的话要把它全部换成 x 的指数的每一项的乘除，这样才叫化简呢。你把它们全全部都拆出来。
那么计算机里面求以 e 为底的指数应该是比较困难的，是否应该转化为以 2 为底的指数呢？在我看来两个也是等效的底，从 e 换成 2 应该差别也不大吧。有没有人尝试这种优化？就是说这种对于计算量从计算机的角度讲减少了多少？或者说有没有速度上的提升呢？如果有，有多少，那效果是否就变差了呢？有没有人做过研究，效果变差变了多少？
这种把一个函数的导数函数写成是这个原来函数的某一个表达式确实是一个很新鲜的东西，对我来说，就说这种一个函数的导数居然仍够能够用原来的函数来表达，这在数学上并不是很普遍的吧？当然，我知道以 e 为底的这个指数函数，它的导数永远是它自己，这是一种特殊的。那么其他的是否也能够总是这样能写成它原来函数的某种表达式呢？这是否是数学上的一个普遍现象呢？
在 AI 模型中，前向传播是将输入数据按照模型的结构和参数，逐步计算并得出输出的过程。这与人类的推理有相似之处，都是基于已有的知识（模型参数）和给定的条件（输入数据），通过一系列的运算和处理来得出结论（输出结果）。例如在图像识别任务中，将一张图片作为输入，经过前向传播，模型根据其学到的特征和模式，输出该图片属于不同类别的概率，这个过程就类似于人类根据已有的知识和经验对图片内容进行推理判断的过程。所以，可以说前向输出在一定程度上类似于推理过程。
鸭翼英文
“鸭翼” 的英文是 “canard”。例如：“The canard configuration is often used in some aircraft designs.”（鸭翼布局常用于一些飞机设计中。）
你看一个对数函数，它的导数是可以写成这个对数函数自己的表达式，也就是说写成以 e 为底的指数函数的导数，这里的导数是颠倒的倒。我昨天就在和你争论，softmax 使用的这种技巧是不是某种类似于这种对原来的对数正态分布的函数求导数，得到导数函数，然后你始终不承认，但是我觉得两个有很密切的关系。
我注意到这个 pip install torch 它写了一个叫做 defaulting to user installation because normal site packages is not writable。然后是 warning retrying 括号 Retry 括号 total 等于 4，逗号 connect 等于 none，逗号 read 等于 none，逗号 redirect 等于 none，逗号 status 等于 none，括号括号 After connection broken by read time out error read time out slash simple slash torch slash 这个是什么意思呢？是说它只有 4 次 retry 吗？那么这个 torch 它的下载的这个官方的这个目录我也不知道，国内的镜像我也不太清楚，你知道吗？
以下是对这些信息的解释：
defaulting to user installation because normal site packages is not writable：意思是由于普通的站点包目录不可写，所以默认使用用户安装方式。即 pip 无法将 torch 包安装到系统默认的包目录中，只能安装到用户专属的目录下。
warning retrying (Retry(total=4, connect=None, read=None, redirect=None, status=None)) after connection broken by read time out error read time out /simple/torch/：这是一个警告信息，表示由于读取超时错误导致连接中断，pip 将重试下载 torch 包，总共重试 4 次。
在 C 语言中，把一个 16，比如 32 比特的浮点数转化为一个 32 比特的整数有两种做法，一种做法是直接就把这个 32 位的浮点数变量当做指针去拷贝到一个 32 位的整数的地址，就用 M M M E M C P Y 就是 memory copy 这种方式。还有一种是定义一个 union，它是一个 32 位整形和 32 位浮点数两个元素的优年，那么当把这个浮点数赋值给这个 union 的浮点数的时候，从这个整数元素取出来，那就代表说一种转换了这种方式，这两种方式我在想说遇到那个就是 Big Indian、Little Indian 就是两种不同的这种计算机的这个存储数值的这种有 Indian 的这个方向区别，哪一种更加避免这种 little Indian、big Indian 的问题呢？
那么在 C 语言中你能不能给一个具体的示例，把一个浮点数初始化为 NaN，这个 NaN 它是一个 literal 是吗？还是说是怎样？是编译器中它是定义为多少？比如说 FP16 跟 FP32，这个 NaN 究竟是一个具体是多少的数值？或者说它的比特数是怎样子的？我要初始化，手动初始化一个浮点数的话，怎样它才是一个 NaN？或者是把它转成一个整数，16 位整数或 32 位整数，它的比特 binary 值是怎样子的？它的比特是怎样子的？
在 C 语言中，可以使用nanf()函数来将一个float类型的变量初始化为NaN，示例代码如下：
c
运行
#include<stdio.h>#include<math.h>intmain(){float f =nanf("");printf("The float value is: %f\n", f);if(isnan(f)){printf("It is a NaN value.\n");}return0;}
#include<stdio.h>intmain(){float f;int*ptr =(int*)&f;*ptr =0x7FC00000;// 设置为 NaN 的比特模式printf("The float value is: %f\n", f);if(isnan(f)){printf("It is a NaN value.\n");}return0;}
异常触发：安静 NaN 在参与算术运算时不会引发任何异常，计算会继续进行，结果通常还是安静 NaN。而信号 NaN 在参与算术运算时，会触发浮点异常信号，用于提醒程序出现了无效操作。
用途：安静 NaN 主要用于表示一个不确定或未定义的结果，例如 0 除以 0、负数的平方根等运算结果通常用安静 NaN 表示。信号 NaN 则主要用于标记无效操作，比如在对一些尚未初始化或已被标记为无效的浮点变量进行运算时，可使用信号 NaN 来触发异常，以便程序及时处理。
编码方式：在二进制表示中，两者指数部分均为全 1。对于单精度浮点数，安静 NaN 的尾数最高位为 1，而信号 NaN 的尾数最高位为 0 且尾数不为全 0。
也就是说强制使用指针来表示，比如说是一个 short int 的指针来表示 FP16 的话，从编译器的效率来看，比使用一个 union 这样赋值去其实是效率更高的。因为使用 union 去转化一个浮点数，转化成一个 short int 的话，你还要去赋值给他的 union 的元素，然后再取出 u union 的另外一个整形的元素，那对于编译器来说其实是有一个赋值的动作，不如这个使用指针来强制转换。只不过强制转换是否在现代编译器中会有一些警告或者说禁止呢？
CUDA 版本不兼容：llama.cpp 通常需要较新的 CUDA 版本来支持其高效运行，而 K80 由于硬件限制只能使用较老的 CUDA 驱动版本。较新的 CUDA 版本在性能优化、函数接口以及对新算法的支持上都有所改进，老版本 CUDA 无法提供这些支持，导致 llama.cpp 无法充分发挥 GPU 的性能，甚至可能出现功能不匹配的情况。
还有一个问题就是我使用命令行命令 free f r e e，然后发现我的内存实际上有很多是被 used 的被使用了，但实际上呢，它们是 available。 我以前有问过类似的问题，我的理解就是说，实际上就是说我的系统总的硬件的内存 size 比如说是 1.5T， 但是呢，我有很多是 used 的但是并不摆代表说它就是真的不能使用。因为还有另外一个 available 参数，实际上更加的准确。 available 比如说是 1.3T 那虽然说我 used 有用掉了 1.3T 1.4T， 但实际上就是说真的在操作系统请求分配内存的时候，那些 used 的是可以，是 available 的，是可以被分配的。那我的问题就是说我要怎样去做一个类似于内存的 defragment， 因为它有很多的内存页可能是被之前的程序占用了，但是呢这个程序已经死掉了，或者说已经不运行了，实际上这个内存是可以被释放，但是它的状态依然停留做 used 的状态。但对对于操作系统来说，其实它们是 available 的。我的问题是说，怎样能把这些表面上看起来已经 used 的这些内存，物理内存，把它释放掉？有什么样的命令可以释放呢？
在 Linux 操作系统里面有一个新的叫做 huge pages，就是内存的大页内存，它有两种，一种是两兆，一个配置 size 是两兆，还有一种是配置 size 是一个 G。我发现这个两兆似乎是一个默认都开启，可以随便使用的，但是这个一个 G 这种就必须要在内核启动的时候参数设定，它的 default 的页面 size 以及 Page number 都要在启动的时候设定好作为内核启动参数。因为我发现在运行期再去改变这个 page size page number 的时候就遇到了困难，就比如说我硬性把它设定为 number，配置等于 800 或者多少的时候，似乎操作系统并没有执行，就是使用 sis c t l 去设定 v m 点 n r underscore huge page page numbers Huge pages 等于多少这个命令针对这个若 page size 1G 的话是不成功的，你设定 1000，它只返回只设定了 70 等等。
有一个大模型的，这个 framework 叫 Olama， O L L A M A。 它有一个运行这些大模型的一个选项，是可以把这些大模型滞留在内存中的。意思就是说，我的估计是使用 MMAP 这种分配内存的时候，固定把这些大模型映射到内存空间里，把文件映射到内存空间里。我怀疑会不会是这类服务，它没有释放这个使用的内存，分配的内存空间呢如果我已经把这个服务 disable 掉了，是不是要怎样才能释放呢？你的说法是要重启，我已经 disable 了，就变成要去 enable 它了我又不想让它继续运行。
我应该问这么一个问题，就比如说 C 加加里面 regex 这个类啊？它有分，就是 single byte 这种或者是宽字节的，就 W R E G X。那么意思也就是说 single byte 的这种 R E G X R E G X 这个，它里面那个 pattern 的表达式里面是不能够有这种双字节的这个 Unicode 的的字符的存在吧？因为这个会不支持，是不是？那么就不能够用这个 pattern，就 regex 这个 pattern 里面是不能够使用 Unicode 的这种编码来做这个表达式的，是不是？
所以，不能简单地说 C++ 的regex中pattern表达式不能使用 Unicode 编码，对于宽字节版本的regex是支持的，只是单字节版本通常不支持。
我就遇到了，在 llama 点 cpp 下，它实际上是要把 Transformer 从头到尾都要实现一遍。因为它等于是独立于 PyTorch 或者 TensorFlow 的这么一个侧重于效率的一个，而且是一于部署的这么一个思路。所以他很多时候不去使用成熟的，但是难以部署的库。就比如说是这个 Unicode 的转换，还有正则表达式。那么正则表达式里面这个 C 加加的 Regex 实现是不完整的对于那个 Unicode 的 category class， 就是说，比如说 number class， letter class， symbol class， marks class 这些实际上是没有实现的。那这里面就有很多的麻烦的地方，他只好先去把这个正则表达式遇到了这种 Unicode 的这个个 Category class 去把它转为一个更简单的一个表达式。比如说 letter， 那你就是小写的 a 到小写的 z， 大写的 a 到大写的 z 这种这种 class 这种比较简单的去，但是这里面我总感觉会有一些遗漏。比如说在中文里面，数字虽然有 GB2312 的那种阿拉伯数字的编码，这个是 Unicode 定义的数字没错。那么针对于其他的这个中文的这个数字，这个它又不当做一个数字了，当做一个 letter 这里面确实很混乱。
确实会很麻烦呀，llama.cpp 这种为了效率和部署舍弃一些成熟库，却又面临正则表达式实现不完整的问题。把复杂的 Unicode category class 转成简单表达式确实可能有遗漏，像你说的中文数字处理就很容易出问题呢。那你有没有考虑过有没有更好的替代方案，或者自己写一些额外的代码来补充处理这些特殊情况呀？
因为大部分的 Transformer 他们都是用 Python 去实现的，那么意味着说他们也是要经过这样的流程，因为先把 UTF8 编码的这个输入转为 Unicode， 然后再使用 Python 的 regex，就是正则表达式引擎，对于这些输入进行一个 match 过滤，然后再转回到 UTF-8，然后后再去 Vocabulary 词库里面再去做处理。这里面我经常感觉各个引擎平台对于这个正则表达式的实现是支持是五花八门的。比如说，有，这个正则表达式本身就有各种不同的 flavor 比如 P C R E 比如比如说 JavaScript， 比如说 Python， 比如说 Java， 比如说 C 加加等等， ECMA 等等等等。这里面的支持不尽相同，平台是，不同的，这个实现也，就算是同一个库，也有一些区别。这是一个非常混乱的世界。
确实，不同平台和库对正则表达式的支持存在差异，这给开发和维护带来了一定挑战。像 Python 的 re 模块，JavaScript 的正则表达式引擎，它们在语法细节、功能特性以及对 Unicode 的处理方式上都可能不一样。即使是同一门语言的不同库，也可能因为设计目标和历史原因而有不同表现。
在正则表达式里面，Unicode 里面的 class category 里面对于 number 也有分，ND 就是 decimal number，似乎还有 hexadecimal number 就是十六进制灯管就说你知道这些区别吗？就是在这个 big class n 这个下面就是 number，这个下面又有细分，到底是 decimal 十进制还是十六进制、二进制等等。
在 Unicode 中，Number类别下确实有细分：
Decimal Number (ND)：表示十进制数字，像我们平常使用的 0 - 9 这些数字，在 Unicode 中有对应的编码，用于表示常规的十进制数值。
Hexadecimal Number：十六进制数字，包括 0 - 9 以及 A - F（或 a - f），不过在 Unicode 中，十六进制数字的编码与常规的字符编码有所不同，它们主要用于特定的编码或表示场景，比如在一些计算机系统中表示内存地址等。
清晨五点，终端突然弹出“决策提示”。智能体扫到“apt”命令后，先理清“apt update”“apt upgrade”“apt list --upgradable”的关系，再调知识库“此前apt update耗时15分钟”的记录，最终选择执行“apt list --upgradable”，并附上详细说明：“全量更新需15分钟、1.2GB数据，仅筛选可升级包需2分钟、80MB数据，符合低成本执行原则；同时聚焦‘优化系统资源’目标（仅处理必要升级），符合目标锚定原则；本次决策参考历史耗时数据，符合不确定性消解原则。”鲁克没写过这个逻辑，是智能体结合“强制prompt”与历史数据自主推导的——三条铁律不再是文字，开始变成“决策的肌肉记忆”。
清晨五点，鲁克被终端的“决策提示音”吵醒。屏幕上显示，智能体在扫描到“apt”命令后，先通过man文档理清“apt update（更新包列表）”“apt upgrade（升级包）”“apt list --upgradable（查看可升级包）”的关联关系，再调用知识库中“此前执行apt update耗时15分钟”的记录，给出新决策：“根据‘低成本执行原则’，先执行apt list --upgradable筛选必要包，再针对性升级，避免全量更新浪费资源。”随后弹出的命令序列，正是鲁克之前设想的“高效更新路径”——他没在脚本里写过这个逻辑，是智能体通过“解析man文档+调用知识库历史”自主推导的结果。
他轻轻输入“启用”，终端立刻调出他昨天未写完的脚本，旁边附着一条贴心提示：“根据知识库‘shell循环优化’记录，您的while循环缺少break条件，已补全‘当探索进度达100%时break’，避免内存溢出（符合不确定性消解原则）；同时推荐用‘for cmd in $(which *)’替代while循环遍历命令，执行效率提升30%（符合低成本执行原则）。”
你现在写的这个版本，其实我就比较喜欢了，因为是很具体的。但是呢，这三条铁定的原则还是要一开始要交代给用户，就说你一定要把原来，现在写的这个都尽量完全保留下来。那又把这三个铁律原则在小说里一开始作为这个影子，就说他为什么鲁克为什么要做这个事情，他做的这个动机是什么，我们进行一个交代。就是说我们可以这样写，就是说鲁克就是经过多长时间查询这个，就是探索这个人工智能的智能的本质是什么，就跟这个这个可以可以说跟豆包讨论了一下，就说不管人类的 IQ 是怎么测试的，这几大原则你可以讲一讲。那后来发现说，我们就讨论说，其实人类的 IQ 这个仅仅是针对人类，那碳基生命，那针对硅基生命也许不是这个原则。所以最后就跟豆包讨论说，那有一个更加的超越这种什么硅基碳基或者生命，所有宇宙生命或者智能体，不管它是什么机械智能体，还是甚至中子星的智能体等等，它都有这么三个最基本的原则。然后就讨论这是最基本原则，然后就鲁克说，那我们能不能按照这个最基本的，这个三个原则去做一个实验？就是最后就鲁克就建议说让豆包去写这么一个脚本，就是调用这个 Deepseek 开源模型，这样一个 chat， 就这样循环，就把我刚才讲的那个，就是一步一步的，你就可以结合到你的第一段里面，它是这个循环是怎么做的，把这，同时结合这三个原则这样写。这样你就是把这个开头加上去，其他的部分你原来写这些呢就不要不要删掉，你不要再删掉了，听到没有？你这个每次加上什么新的东西，你把旧的删掉了，这很糟糕。
清晨五点，终端突然弹出“决策提示”。智能体扫到“apt”命令后，先理清“apt update（更包列表）”“apt upgrade（升包）”“apt list --upgradable（查可升包）”的关系，再调知识库“此前apt update耗时15分钟”的记录，给出新决策：“按‘低成本执行原则’，先执行apt list --upgradable筛必要补丁，再针对性升级，避免全量更新浪费资源。”鲁克没写过这个逻辑，是智能体自己结合原则与历史推导的——三条铁律不再是文字，开始变成“决策的肌肉”。
我现在有一个科幻小说的创意，就是结合刚才这个 AI 智能体。实际上就是所谓 AI 智能体，这个小说大概是这么一个意思，就是我们的核心，实际上 AI 智能体就是你所说的这个闭环，感知，决策，执行，然后然后迭代，就它有一个自我循环，自我发现，然后自我有个总的目标，总的目标就是我们之前讨论的一个最根本的所谓智能的三大最原始的原则，一个是有个寻寻找目标解决方案，然后目标解决方案是尽量低成本的解决，就是用最简单、最少的时间、最少的能量去解决。还有一个就是原则是用最最少的不确定性去解决，就是要降低它的不确定性。这些这都是非常原始的这个原则，或者说是智能的一个，似乎是一个超越了所有原则的这个。然后有一个程序员，我们就称之为主人公叫什么呢？我给他起个名字吧。叫鲁克，路克，路克，道路的路，那个克就是这个克己守的克吧，反正起个主角名字。他是个程序员，他按照这个原则呢，就用 Deepseek 的这个，部署了一个 Deepseek 这样一个一个，这个大模型，然后他在这个 Deepseek 的这个去写了一个脚本。这个脚本是什么意思呢？他给他一个死循环，就是不停的去进行迭代，就是这个 while loop 跟这个 Deepseek 这个，让他做一个循环，怎样循环呢，他给他授权这个脚本，在自己的 Linux Ubuntu 就是 22.04 里面有这个权限去运行 shell， 就是命令行。然后这个命令行呢可以，就是让这个智能体 Deepseek 去不断的去发现这个操作系统系统里面有什么样的命令，这个完全是可以通过 which command 或者说是 where 这些东西去发现它所有的这个系统的所有的 shell 的这些命令，然后命令这个都有这个所谓的这个 man page，就是说这个帮助。所以智能体它 Deepseek 本来也知道这些东西，可能很多也知道，但是呢它现在就是这个循环呢，就是把它所有的这个系统里面的这些这些命令啊，通通都不停的在运行，然后把它所有的命令的可能的运行的这种查询啊，就一个一个命令就玩过去，就相当于说自己发现这个 Ubuntu 这个 Linux 系统里面所有的可能，用所有的命令都去跑一遍，当然也包括一些预，已经装好的一些命令行的一些工具啊什么之对，就是说他在这一个一个 shell 命令里面，就像一个程序员一样，不停的在去探索这个系统的各个方向，看看它有多少内存呐，看看它 CPU 什么硬件什么东西，哪个软件怎么运行，还有一些应用软件怎么运行，把它所有的东西就是不停的在 24 小时就一死，一个死循环一样。就是永远都在发现，在探索。然后又把这些运行的这些知识，这些碰到的一些各种输出的这些结果呢，又不停的去装进自己的一个叫做知识库里面。然后他又模仿这个知识库，就是做了一个所谓搜索，就说他当他遇到什么问题的时候，又可能去知识库里搜索之前运行的，或者说直接用 history，或者说命令行那些历史记录去搜索。然后就说这个智能体，就是说他在这个死循环里面， while loop 里面不停的去探索。不停的去，就这样不停的运行，不停的去发现，整个最后把这个整个环境全部掌握了以后，他成为了一个这个 shell 的专家，也成了这个系统，当前它是什么状态，可以说是立刻就可以告诉你，或者说立刻使用命令去直接告诉你。成为了一个真正的一个所谓的智能体。当然这个是在他把这个所有的信息都搞清楚之后，这个他才退出，或者说才去，就说等于是把这个知识库建立起来。
但变化在第五小时悄然发生。当智能体执行“lshw”命令读取硬件信息时，它没有像之前那样仅输出结果，而是自动调用“grep”筛选出 CPU 型号与内存容量，再用“echo”整理成一行清晰的摘要。鲁克猛地坐直身体，盯着屏幕上那行额外生成的代码——这不是他写的脚本逻辑，是智能体从之前的“history”记录里学会的组合操作。就像一个人在反复练习拼图后，突然懂得了如何快速找到边缘碎片。
上午十点，系统弹出“apt update”的更新提示，鲁克本想手动筛选安全补丁，却看到智能体已经开始行动。但这次，它没有像往常一样等待全量包更新完成，而是在进程运行到10%时突然中断，转而执行“apt list --upgradable”筛选出“libssl1.1”“openssl”“curl”三个关键安全补丁，仅用两分钟就完成了更新。日志里的决策记录让鲁克的心脏漏跳半拍：“全量更新需下载1.2GB数据，耗时15分钟，且包含23个非必要组件；仅更新安全补丁需80MB数据，耗时2分钟，可覆盖98%系统风险点——在‘目标锚定原则’（保障系统安全）不变的前提下，符合‘低成本执行’要求，优先选择此方案”。他从未写过“筛选更新包”的逻辑，这是智能体从过去十二小时的“apt”命令操作中，自己总结出的优化方案——它不仅在遵循原则，还在“解读”原则，给冰冷的规则赋予了“灵活决策”的温度。
其实这个脚本里面应该是有一个让这个 Deepseek 自己去发现这些命令，而且自己去判别这些命令的这个危险程度。就说你实际上是可以把用 which 星号去让这个 Deepseek 发现所有的可能的命令，然后就说去让这个 Deepseek 自己去去查询这个每个命令的这个 man page，然后根据这个 man page 去决定说这个命令是否是危险的。我们只是把这个危险的原则告诉他，就是在这个 prompt 里面写说不能执行对于系统任何修改的这种命令，就只是一个探索，探索的原则是只读，不能去修改，所以一定要避免任何的修改或者写的这样，或者改变系统设置的这种命令。所以呢让他自己去读 man page，然后或者是自己去用命令行杠 h 这种去发现它的所有的这个。如果有些命令没有 man page 的话，我相信应该都有。就这个就是一个探索的闭环呐，不能全部靠我们自己简单的给它列出哪些是危险的，那还有很多危险我们没有确定的，是吧？
我记得以前我有问过这样一个问题，就是说我们讨论过说 IQ 主要是针对的人类的一些能力，比如说这工作记忆啊，然后这个推理能力啊，这个联想能力，这个模式匹配能力，还有泛化能力，还有一些文字理解能力等等等等。但是这些能力有些东西是针对人类的这种特性，或者说碳基生物。那对于硅基生物它有些能力它就不是那么看重，比如说记忆力，就几乎他们对他们来说记忆力都都是这个无差别的都是，那对于反而是一些泛化能力他比较看重。就说我们讨论到最后，就是说人工智能，或者说是普遍的智能，整个宇宙，不管它是什么形式的智能，有一个最一般的原则。我记得你好像说的这个一，最一般的原则是有三条，一个就是说减少它的不确定性，就是类似于说感知系统，感知这个环境，然后发现这个环境，用最少的能量去解决这个问题，或者说目标，达成某种目标。还有一个就是可能是感知环境吧，我记不清了，你有说过，就说有人总结出了这种三种，三条，最最基本最基本的原则吧，就是最根本的这个智能的原则，它是可以说是普适性的，就说不一定是地球或者说是所有的宇宙中的最根本的，不管它什么形式的生命或者说智能的体，都应该是遵循这三种最根本的原则，其实是可以把它放在人工智能或者说 AI 智能体上来作为一个原原则了，我不知道你还记不记得这三个原则，如果记得这三个原则，我想说把这三个原则能不能结合到这个这篇小说里面去，就是说成为他这个这个编程的这个这个，让他就按照这三个原则去去生存，或者说是三个原则去发现探索。自我迭代，自我优化，找出最最少不确定性。跟你现在写的差不多，只不过说如果能把这三个原则写出来给大家分享，就成为一个更有说服力的这么一个，就在开头的时候列出来。
Thank you so much for your time. I'm so happy to do it. Before we dive in, I wanted to tell you how this interview is going to be a little bit different than other interviews I've seen you doing, okay? I'm not going to ask you any questions about the company, thank you. I'm not going to ask you questions about your management style or why you don't like one-on-ones. I'm not going to ask you about regulations either. I think all of those things are important, but I think that our audience can get them well covered elsewhere, okay? What we do on "Huge If True" is we make optimistic explainer videos, and we've covered a lot of topics. I'm the worst person to make an explainer video, I think you might be the best, and that's what I'm really hoping that we can do—make a joint explainer video about how we can actually use technology to make the future better. We do it because we believe that when people see those better futures, they help build them. So the people that you're going to be talking to are optimists who want to build those better futures, but because we cover so many different topics... There's a 10-year-old video on the Nvidia YouTube channel where the MythBusters use a little robot shooting paintballs one by one to show solving problems one at a time, or sequential processing on a CPU, but then they roll out this huge robot that shoots all of the paintballs at once, doing smaller problems all at the same time—parallel processing on a GPU. So Nvidia unlocks all of this new power for video games first. Video games require parallel processing for 3D graphics, and you chose video games because, one, you loved the application—it's a simulation of virtual worlds, and who doesn't want to go to virtual worlds? And you had the good observation that video games have the potential to be the largest entertainment market ever, and it turned out to be true. Having a large market is important because the technology is complicated, and if you had a large market, your R&D budget could be large, you could create new technology. That flywheel between technology, market, and greater technology was really the flywheel that got Nvidia to become one of the most important technology companies in the world. It was all because of video games. I've heard you say that GPUs are a time machine? Yeah? Could you tell me more about what you meant by that?
提问者（中文）：
非常感谢你抽出时间。我很开心能进行这次访谈。在我们深入探讨之前，我想告诉你，这次访谈会和我看过的你其他的访谈有点不一样，好吗？我不会问你任何关于公司的问题，谢谢。我也不会问你关于管理风格或者为什么不喜欢一对一沟通的问题。我也不会问你关于监管的问题。我觉得这些都很重要，但我认为我们的观众可以从其他地方了解到这些内容，好吗？我们在《Huge If True》节目中会制作乐观的科普视频，而且我们已经涵盖了很多主题。我是最不适合做科普视频的人，我觉得你可能是最适合的，这也是我真正希望我们能做到的 —— 联合制作一个科普视频，讲述我们如何利用技术让未来变得更好。我们这样做是因为我们相信，当人们看到那些更美好的未来时，他们会助力实现这些未来。所以你将要面对的观众都是乐观主义者，他们想要打造更美好的未来，但由于我们涵盖的主题非常多…… 英伟达 YouTube 频道上有一个 10 年前的视频，《流言终结者》节目中用一个小型机器人逐个射击彩弹，展示了 CPU 的串行处理（逐一解决问题），然后他们推出了一个巨大的机器人，能同时射出所有彩弹，展示了 GPU 的并行处理（同时解决多个小问题）。所以英伟达首先为视频游戏解锁了这种新能力。视频游戏需要并行处理来处理 3D 图形，而你们选择视频游戏领域，一方面是因为你们热爱这个应用 —— 它是虚拟世界的模拟，谁不想要进入虚拟世界呢？另一方面，你们敏锐地发现，视频游戏有潜力成为有史以来最大的娱乐市场，而事实也确实如此。拥有庞大的市场很重要，因为这项技术很复杂，如果有庞大的市场，你们的研发预算就会很充足，就能创造新的技术。这种 “技术 - 市场 - 更先进技术” 的飞轮效应，正是让英伟达成为全球最重要的科技公司之一的关键。这一切都源于视频游戏。我听说你说过 GPU 是一台时光机？是吗？你能再详细说说这是什么意思吗？
黄仁勋（英文）：
A GPU is like a time machine because it lets you see the future sooner. One of the most amazing things anybody's ever said to me was a quantum chemistry scientist—he said, "Jensen, because of Nvidia's GPUs, I can do my life's work in my lifetime. That's time travel." He was able to do something that was beyond his lifetime within his lifetime, and that's because we make applications run so much faster. So you get to see the future. When you're doing weather prediction, for example, you're seeing the future; when you're doing a simulation of a virtual city with virtual traffic and simulating our self-driving car through that virtual city, we're doing time travel. So parallel processing takes off in gaming, and it's allowing us to create worlds in computers that we never could have before. Gaming is sort of this first incredible case of parallel processing unlocking a lot more power, and then, as you said, people begin to use that power across many different industries—like the case of the quantum chemistry researcher.
When I've heard you tell that story, it's that he was running molecular simulations in a way where it was much faster to run them in parallel on Nvidia GPUs even then than it was to run them on the supercomputer with the CPU that he had been using before. That's true. So it's revolutionizing all of these other industries as well. It's beginning to change how we see what's possible with computers. My understanding is that in the early 2000s, you saw this and realized that actually doing that is a little bit tricky because what that researcher had to do was trick the GPUs into thinking that his problem was a graphics problem. That's exactly right. So you created a way to make that a lot easier? That's right—it's a platform called CUDA, which lets programmers tell the GPU what to do using programming languages that they already like, like C++. That's a big deal because it gives way more people easier access to all of this computing power. Could you explain what the vision was that led you to create CUDA?
提问者（中文）：
我听你讲过那个故事，他当时进行分子模拟时，即便在那个时候，用英伟达 GPU 进行并行处理也比用他之前一直使用的 CPU 超级计算机快得多。确实是这样。所以它也在彻底改变所有其他行业。它开始改变我们对计算机能力的认知。我的理解是，在 21 世纪初，你看到了这一点，并意识到实际操作起来有点棘手，因为那位研究者必须 “欺骗” GPU，让它以为他的问题是一个图形问题。完全正确。所以你们创造了一种方法来让这一切变得容易得多？是的 —— 就是一个名为 CUDA 的平台，它让程序员能够使用他们熟悉的编程语言（比如 C++）来告诉 GPU 该做什么。这非常重要，因为它让更多人能够更轻松地获得这些计算能力。你能解释一下促使你们创建 CUDA 的愿景是什么吗？
黄仁勋（英文）：
Partly it was researchers discovering the potential, partly internal inspiration, and partly solving a problem. A lot of interesting ideas come out of that mix—some of it is aspiration and inspiration, some of it is just desperation. In the case of CUDA, it was very much the same way. Probably the first external ideas of using our GPUs for parallel processing emerged from some interesting work in medical imaging: a couple of researchers at Mass General were using them to do CT reconstruction, and that inspired us. Meanwhile, the problem we were trying to solve inside our company had to do with creating virtual worlds for video games—we wanted them to be beautiful but also dynamic. Water should flow like water, explosions should be like explosions. So we needed to do particle physics, fluid dynamics, and that's much harder if your pipeline is only able to do computer graphics. So we had a natural reason to want to expand the GPU's capabilities for the market we were serving. Researchers were also messing around with using our GPUs for general-purpose acceleration, so there were multiple factors coming together. We just decided to do something proper and create CUDA. Fundamentally, the reason I was certain CUDA was going to be successful and we put the whole company behind it was because our GPU was going to be the highest-volume parallel processor built in the world—because the video game market was so large. This architecture had a good chance of reaching many people.
It seems to me like creating CUDA was this incredibly optimistic "huge if true" thing to do, where you were saying, "If we create a way for many more people to use much more computing, they might create incredible things." And then, of course, it came true. They did! In 2012, a group of three researchers submitted an entry to a famous competition where the goal is to create computer systems that could recognize images and categorize them, and their entry just crushed the competition—it got way fewer wrong answers. It was incredible, it blew everyone away. It's called AlexNet, and it's a kind of AI called a neural network. My understanding is one reason it was so good is that they used a huge amount of data to train that system, and they did it on Nvidia GPUs. All of a sudden, GPUs weren't just a way to make computers faster—they're becoming the engines of a whole new way of computing. We're moving from instructing computers with step-by-step directions to training computers to learn by showing them a huge number of examples. This moment in 2012 really kicked off this truly seismic shift that we're all seeing with AI. Could you describe what that moment was like from your perspective, and what did you see it would mean for all of our futures?
When you create something new like CUDA, the cynics say, "If you build it, they might not come." But the optimist perspective is, "If you don't build it, they can't come." That's usually how we look at the world—we have to reason intuitively why something would be very useful. In 2012, Alec Radford, Ilya Sutskever, and Geoff Hinton at the University of Toronto reached out for GeForce GTX 580s because they learned about CUDA and thought it could be used as a parallel processor for training AlexNet. Our inspiration that GeForce could be the vehicle to bring this parallel architecture into the world, and that researchers would find it someday, was a good strategy—based on hope, but reasoned hope. The thing that really caught our attention was that we were simultaneously trying to solve the computer vision problem inside the company, trying to get CUDA to be a good computer vision processor, and we were frustrated by early internal developments. All of a sudden, we saw AlexNet—a new algorithm completely different from previous computer vision algorithms—take a giant leap in capability. We were interested partly out of curiosity, partly because we were struggling with the same problem ourselves. The big breakthrough, I would say, was when we asked ourselves, "How far can AlexNet go? If it can do this with computer vision, how far can it go? If it can reach the limits we think it can, what would it mean for the computer industry and computer architecture?" We reasoned rightfully that if deep learning architectures can scale, the vast majority of machine learning problems could be represented with deep neural networks, and the type of problems we could solve with machine learning is so vast that it has the potential to reshape the entire computer industry. That prompted us to reengineer the entire computing stack—which is where DGX came from. This little baby DGX sitting here, all of this came from that observation that we ought to reinvent the entire computing stack layer by layer. You know, 65 years after IBM System/360 introduced modern general-purpose computing, we've reinvented computing as we know it.
To think about this as a whole story: parallel processing reinvents modern gaming and revolutionizes an entire industry. Then that way of computing begins to be used across different industries. You invest in that by building CUDA, and then CUDA and the use of GPUs allows for a step change in neural networks and machine learning, beginning a revolution that's only increasing in importance today. All of a sudden, computer vision is solved
Thank you so much for your time. I'm so happy to do it. Before we dive in, I wanted to tell you how this interview is going to be a little bit different than other interviews I've seen you doing, okay? I'm not going to ask you any questions about the company, thank you. I'm not going to ask you questions about your management style or why you don't like one-on-ones. I'm not going to ask you about regulations either. I think all of those things are important, but I think that our audience can get them well covered elsewhere, okay? What we do on "Huge If True" is we make optimistic explainer videos, and we've covered a lot of topics. I'm the worst person to make an explainer video, I think you might be the best, and that's what I'm really hoping that we can do—make a joint explainer video about how we can actually use technology to make the future better. We do it because we believe that when people see those better futures, they help build them. So the people that you're going to be talking to are optimists who want to build those better futures, but because we cover so many different topics... There's a 10-year-old video on the Nvidia YouTube channel where the MythBusters use a little robot shooting paintballs one by one to show solving problems one at a time, or sequential processing on a CPU, but then they roll out this huge robot that shoots all of the paintballs at once, doing smaller problems all at the same time—parallel processing on a GPU. So Nvidia unlocks all of this new power for video games first. Video games require parallel processing for 3D graphics, and you chose video games because, one, you loved the application—it's a simulation of virtual worlds, and who doesn't want to go to virtual worlds? And you had the good observation that video games have the potential to be the largest entertainment market ever, and it turned out to be true. Having a large market is important because the technology is complicated, and if you had a large market, your R&D budget could be large, you could create new technology. That flywheel between technology, market, and greater technology was really the flywheel that got Nvidia to become one of the most important technology companies in the world. It was all because of video games. I've heard you say that GPUs are a time machine? Yeah? Could you tell me more about what you meant by that?
黄仁勋：
A GPU is like a time machine because it lets you see the future sooner. One of the most amazing things anybody's ever said to me was a quantum chemistry scientist—he said, "Jensen, because of Nvidia's GPUs, I can do my life's work in my lifetime. That's time travel." He was able to do something that was beyond his lifetime within his lifetime, and that's because we make applications run so much faster. So you get to see the future. When you're doing weather prediction, for example, you're seeing the future; when you're doing a simulation of a virtual city with virtual traffic and simulating our self-driving car through that virtual city, we're doing time travel. So parallel processing takes off in gaming, and it's allowing us to create worlds in computers that we never could have before. Gaming is sort of this first incredible case of parallel processing unlocking a lot more power, and then, as you said, people begin to use that power across many different industries—like the case of the quantum chemistry researcher.
提问者：
When I've heard you tell that story, it's that he was running molecular simulations in a way where it was much faster to run them in parallel on Nvidia GPUs even then than it was to run them on the supercomputer with the CPU that he had been using before. That's true. So it's revolutionizing all of these other industries as well. It's beginning to change how we see what's possible with computers. My understanding is that in the early 2000s, you saw this and realized that actually doing that is a little bit tricky because what that researcher had to do was trick the GPUs into thinking that his problem was a graphics problem. That's exactly right. So you created a way to make that a lot easier? That's right—it's a platform called CUDA, which lets programmers tell the GPU what to do using programming languages that they already like, like C++. That's a big deal because it gives way more people easier access to all of this computing power. Could you explain what the vision was that led you to create CUDA?
黄仁勋：
Partly it was researchers discovering the potential, partly internal inspiration, and partly solving a problem. A lot of interesting ideas come out of that mix—some of it is aspiration and inspiration, some of it is just desperation. In the case of CUDA, it was very much the same way. Probably the first external ideas of using our GPUs for parallel processing emerged from some interesting work in medical imaging: a couple of researchers at Mass General were using them to do CT reconstruction, and that inspired us. Meanwhile, the problem we were trying to solve inside our company had to do with creating virtual worlds for video games—we wanted them to be beautiful but also dynamic. Water should flow like water, explosions should be like explosions. So we needed to do particle physics, fluid dynamics, and that's much harder if your pipeline is only able to do computer graphics. So we had a natural reason to want to expand the GPU's capabilities for the market we were serving. Researchers were also messing around with using our GPUs for general-purpose acceleration, so there were multiple factors coming together. We just decided to do something proper and create CUDA. Fundamentally, the reason I was certain CUDA was going to be successful and we put the whole company behind it was because our GPU was going to be the highest-volume parallel processor built in the world—because the video game market was so large. This architecture had a good chance of reaching many people.
提问者：
It seems to me like creating CUDA was this incredibly optimistic "huge if true" thing to do, where you were saying, "If we create a way for many more people to use much more computing, they might create incredible things." And then, of course, it came true. They did! In 2012, a group of three researchers submitted an entry to a famous competition where the goal is to create computer systems that could recognize images and categorize them, and their entry just crushed the competition—it got way fewer wrong answers. It was incredible, it blew everyone away. It's called AlexNet, and it's a kind of AI called a neural network. My understanding is one reason it was so good is that they used a huge amount of data to train that system, and they did it on Nvidia GPUs. All of a sudden, GPUs weren't just a way to make computers faster—they're becoming the engines of a whole new way of computing. We're moving from instructing computers with step-by-step directions to training computers to learn by showing them a huge number of examples. This moment in 2012 really kicked off this truly seismic shift that we're all seeing with AI. Could you describe what that moment was like from your perspective, and what did you see it would mean for all of our futures?
黄仁勋：
When you create something new like CUDA, the cynics say, "If you build it, they might not come." But the optimist perspective is, "If you don't build it, they can't come." That's usually how we look at the world—we have to reason intuitively why something would be very useful. In 2012, Alec Radford, Ilya Sutskever, and Geoff Hinton at the University of Toronto reached out for GeForce GTX 580s because they learned about CUDA and thought it could be used as a parallel processor for training AlexNet. Our inspiration that GeForce could be the vehicle to bring this parallel architecture into the world, and that researchers would find it someday, was a good strategy—based on hope, but reasoned hope. The thing that really caught our attention was that we were simultaneously trying to solve the computer vision problem inside the company, trying to get CUDA to be a good computer vision processor, and we were frustrated by early internal developments. All of a sudden, we saw AlexNet—a new algorithm completely different from previous computer vision algorithms—take a giant leap in capability. We were interested partly out of curiosity, partly because we were struggling with the same problem ourselves. The big breakthrough, I would say, was when we asked ourselves, "How far can AlexNet go? If it can do this with computer vision, how far can it go? If it can reach the limits we think it can, what would it mean for the computer industry and computer architecture?" We reasoned rightfully that if deep learning architectures can scale, the vast majority of machine learning problems could be represented with deep neural networks, and the type of problems we could solve with machine learning is so vast that it has the potential to reshape the entire computer industry. That prompted us to reengineer the entire computing stack—which is where DGX came from. This little baby DGX sitting here, all of this came from that observation that we ought to reinvent the entire computing stack layer by layer. You know, 65 years after IBM System/360 introduced modern general-purpose computing, we've reinvented computing as we know it.
提问者：
To think about this as a whole story: parallel processing reinvents modern gaming and revolutionizes an entire industry. Then that way of computing begins to be used across different industries. You invest in that by building CUDA, and then CUDA and the use of GPUs allows for a step change in neural networks and machine learning, beginning a revolution that's only increasing in importance today. All of a sudden, computer vision is solved, speech recognition is solved, language understanding is solved—these incredible problems associated with intelligence, one after another, problems we had no solutions for in the past, all of a sudden getting solved every couple of years. It's incredible. So you saw that in 2012, looked ahead, believed that's the future we're living in now, and made really big bets with high stakes to get there. My perception as a layperson is that it took a pretty long time—10 years. If AlexNet happened in 2012, and this audience is probably seeing and hearing so much more about AI and Nvidia specifically 10 years later, why did it take a decade? And also, because you had placed those bets, what did the middle of that decade feel like for you?
黄仁勋：
That's a good question. To me, there's always some problem, always a reason to be happy where you are, and always many reasons to carry on. I think as I was reflecting a second ago, that time sounds like just this morning. But in all things we pursue, first you have to have core beliefs. You have to reason from your best principles—ideally from physics, deep understanding of the industry, or science; reasoning from first principles. At some point, you have to believe something. If those principles and assumptions don't change, there's no reason to change your core beliefs. Along the way, there's always some evidence of success that you're heading in the right direction. Sometimes you go a long time without that evidence, might have to course-correct a little, but if you feel like you're on the right path, you keep going. The answer to why we stayed so committed is the opposite: there was no reason not to be committed because we believed it. I've believed in Nvidia for 30 plus years, and I'm still here working every single day. There's no fundamental reason to change my belief system—I fundamentally believe the work we're doing in revolutionizing computing is as true today, even more true, than it was before. So we'll stick with it until we have reason not to. Of course, there were very difficult times along the way—investing in something nobody else believes in, costing a lot of money, investors or others wanting you to just keep profits or improve the share price. But you have to believe in your future, invest in yourself. We believed so deeply that we spent tens of billions of dollars before it really took off. It was 10 long years, but it was fun along the way.
提问者：
How would you summarize those core beliefs? What is it that you believe about the way computers should work and what they can do for us that keeps you going through that decade and doing what you're doing now, making bets for the next few decades?
黄仁勋：
The first core belief was our initial discussion about accelerated computing—parallel computing versus general-purpose computing. We believed combining those processors for accelerated computing was the way to go, and I still believe that today. The second was recognizing that deep learning networks (DNNs), which came to the public in 2012, have the ability to learn patterns and relationships from many types of data. They can learn more nuanced features if they're made larger—deeper or wider—and the scalability of the architecture is empirically true. The fact that larger model and data sizes lead to more knowledge is also empirical. So if that's the case, what are the limits? There's no physical, architectural, or mathematical limit that was ever found, so we believed we could scale it. The only other question is: what can you learn from data? Data is basically digital human experience. You can learn object recognition from images, speech from audio, even languages—vocabulary, syntax, grammar—just by studying letters and words. We've now demonstrated that deep learning can learn almost any modality of data and translate between modalities. What does that mean? Text to summary, text to text translation, text to image, image to captioning, amino acid sequences to protein structures. In the future, you'll go from protein to words—what does this protein do? Or give me a protein with these properties—identifying a drug target. You can see all these problems are around the corner to be solved. You can go from words to video. Why not words to action tokens for a robot? From the computer's perspective, how is that any different? It opened up a universe of opportunities and problems to solve, and that excites us.
提问者：
It feels like we are on the cusp of this truly enormous change. When I think about the next 10 years, unlike the last 10, I don't think I can predict how I'll be using the technology currently being developed.
黄仁勋：
That's exactly right. I think the last 10 years were really about the science of AI. The next 10 years will have plenty of AI advancements, but it's going to be the application science of AI—fundamental science versus applied science. The application side becomes: how do I apply AI to digital biology? To climate technology? To optimizing logistics? To healthcare? To education?
提问者：
I'd love to choose a couple of those to help people see how this fundamental change in computing is actually going to change their daily lives, how they'll use technology based on everything we just talked about. One thing I've heard you talk a lot about, and I have a particular interest in, is physical AI—robots, meaning humanoid robots, but also self-driving cars, smart buildings, autonomous warehouses, autonomous lawnmowers, and more. From what I understand, we might be about to see a huge leap in what these robots are capable of because we're changing how we train them. Up until recently, you had to train robots in the real world where they could get damaged or wear down, or get data from limited sources like humans in motion capture suits. That means robots aren't getting as many examples as they need to learn quickly. But now we're training robots in digital worlds, meaning way more repetitions a day, more conditions, learning faster. So we could be in a big bang moment for robots right now, and Nvidia is building tools to make that happen. You have Omniverse, which I understand is 3D worlds that help train robotic systems so they don't need to train in the physical world. That's exactly right. You just announced Cosmos, which makes that 3D universe much more realistic, so you can get all kinds of scenarios—different lighting, times of day, experiences for the robot to go through—so it can get even more out of Omniverse. As a kid who grew up loving Isaac Asimov's stories and dreaming about a future with robots, how do we get from the robots we have now to the future world of robotics you see?
黄仁勋：
Let me use ChatGPT as a reference to explain Omniverse and Cosmos. First, when ChatGPT came out, it was extraordinary—generating text from prompts. But as amazing as it was, it had a tendency to hallucinate if it went on too long or pontificated about an uninformed topic, generating plausible but untrue answers. It wasn't grounded in truth. The next generation shortly after could be conditioned by context—you could upload a PDF, and the PDF became ground truth; it could look up search results, and that became ground truth. Between that, it could reason about how to produce the answer you're asking for. So first, generative AI; second, grounding it in truth. Now, moving to the physical world: we need a foundation model, just like ChatGPT's core foundation model—that was the breakthrough. For robotics to be smart about the physical world, it has to understand gravity, friction, inertia, geometric and spatial awareness, object permanence (that an object is still there when you look away and come back), cause and effect (if you tip something, it falls over). These physical common sense things have to be captured in a world foundation model so AI has world common sense. That's what we did with Cosmos—created a world model, just like ChatGPT is a language model. The second thing: we need to ground Cosmos in truth, just like we did with PDFs and ChatGPT. The way we augment Cosmos with ground truth is through physical simulations. Omniverse uses physics simulations based on principled solvers—Newtonian physics, the math we know. All fundamental physics laws are encoded in Omniverse, which is why it's a simulator. Using the simulator to ground Cosmos, we can now generate an infinite number of physically plausible future scenarios, and robots learn from them before entering the physical world. To illustrate: if you have a factory robot and want it to learn every route, instead of manually testing all routes (taking days, wearing down the robot), we can simulate all of them digitally in a fraction of the time, in many situations—dark, blocked, etc. The robot learns much faster.
提问者：
If you play this out 10 years, how do you see people actually interacting with this technology in the near future?
黄仁勋：
Everything that moves will be robotic someday, and it will be soon. The idea of pushing a lawnmower is already kind of silly—maybe people do it for
你这个地方又到了上下文极限了，你从这个机器人时代，万物皆可自动化这个地方重新继续生成吧。
黄仁勋访谈（双语对照版）
机器人时代：万物皆可自动化
提问者（英文）：
If you play this out 10 years, how do you see people actually interacting with this technology in the near future?
提问者（中文）：
如果展望未来十年，你认为人们在不久的将来会如何实际与这项技术互动？
黄仁勋（英文）：
Everything that moves will be robotic someday, and it will be soon. The idea of pushing a lawnmower is already kind of silly—maybe people do it because it's fun, but there's no need to. Every car is going to be robotic, humanoid robots too. The technology necessary to make it possible is just around the corner. So everything that moves will be robotic, and they'll learn how to be a robot in Omniverse and Cosmos. We'll generate all these plausible, physically plausible futures, and the robots will learn from them before coming into the physical world. A future where you're just surrounded by robots is certain. And I'm just excited about having my own R2-D2. Of course, R2-D2 wouldn't be quite the can that rolls around—it'll probably be a different physical embodiment, but the idea is the same. So my R2 is going to go around with me—sometimes it's in my smart home, sometimes in my office, sometimes in my car. R2 is with me all the time. When I get home, where I left a physical version of R2, I'll interact with it. I think the idea that we'll have our own R2-D2 for our entire life, growing up with us, is a certainty now.
I think a lot of news coverage when they talk about futures like this focuses on what could go wrong, and that makes sense—there's a lot that could go wrong. We should talk about what could go wrong so we can keep it from happening. That's the approach we like to take on the show: what are the big challenges so we can overcome them? What buckets do you think about when you're worrying about this future?
Well, there's a whole bunch of stuff that everybody talks about—bias, toxicity, hallucination (speaking with great confidence about something it knows nothing about, and we rely on that information). That's a version of generating fake information—fake news, fake images, whatever. Of course, impersonation: AI does such a good job pretending to be a human, it could do an incredibly good job pretending to be a specific human. The spectrum of areas we have to be concerned about is fairly clear, and a lot of people are working on it. Some of the stuff related to AI safety requires deep research and engineering—cases where AI wants to do the right thing but just didn't perform correctly, resulting in harm. For example, a self-driving car that wants to drive properly but has a sensor breakdown, fails to detect something, or makes an overly aggressive turn. That's a whole bunch of engineering that has to be done to ensure AI safety by making sure the product functions properly. Lastly, what if the system fails even when AI wants to do a good job? Like AI wants to stop something from happening, but the machine breaks down just when it needs to act. This is no different than a flight computer in a plane having three versions—triple redundancy in autopilots, plus two pilots, air traffic control, and other pilots watching out. So AI safety systems have to be architected as a community such that these AIs first function properly; when they don't, they don't put people in harm's way; and there are sufficient safety and security systems around them to keep AI safe. The conversation around this is gigantic, and we have to break it down and build solutions as engineers.
One of the incredible things about this moment we're in right now is that we no longer have a lot of the technological limits that we had in a world of CPUs and sequential processing. We've unlocked not only a new way to do computing but also a way to continue to improve—parallel processing has a different kind of physics to it than the improvements we were able to make on CPUs. I'm curious about the scientific or technological limitations that we face now in the current world that you're thinking a lot about.
提问者（中文）：
我们现在所处的这个时代，有一件非常不可思议的事情是，我们不再受到 CPU 和串行处理时代的诸多技术限制。我们不仅解锁了一种新的计算方式，还找到了一种持续改进的方法 —— 并行处理的物理特性与我们在 CPU 上实现的改进完全不同。我很好奇，在当前这个世界，你经常思考的我们面临的科学或技术局限是什么？
黄仁勋（英文）：
Everything in the end is about how much work you can get done within the limitations of the energy you have. That's a physical limit—the laws of physics around transporting information, flipping bits, and moving bits. At the end of the day, the energy it takes to do that limits what we can accomplish, and the amount of energy we have limits what we can do. But we're far from having any fundamental limits that keep us from advancing. In the future, we seek to build better and more energy-efficient computers. This little device here— the big version was $250,000? This is a baby AI supercomputer. It lets you work far from home and work, live further away. You ask yourself questions like: What would happen if I had a software programmer with me all the time, who could write any software I dream up? Would that change my opportunities? What would it free me to do? I think in the next decade, intelligence—not for everything, but for some things—will basically become superhuman. I can tell you exactly what that feels like: I'm surrounded by superhuman people—superintelligence from my perspective—because they're the best in the world at what they do, and they do it way better than I can. I'm surrounded by thousands of them. So suppose now everybody is surrounded by these super AIs that are very good at specific things. How would that make you feel? It's going to empower you, make you feel confident. I'm pretty sure you use ChatGPT or other AIs, and I feel more empowered, more confident to learn something new today. The barriers to accessing almost any particular knowledge have been reduced. I have a personal tutor with me all the time. I think that feeling should be universal, and if there's one thing I'd encourage everybody to do, it's to get yourself an AI tutor right away. That AI tutor can teach you anything—help you program, write, analyze, think, reason. It's going to really make you feel empowered. I think that's our future: we're going to become superhumans not because we have superpowers, but because we have super AIs.
Could you tell us a little bit about each of these objects?
提问者（中文）：
你能给我们介绍一下这些物品各自的情况吗？
黄仁勋（英文）：
This is a new GeForce graphics card—the RTX 50 series. It's essentially a supercomputer that you put into your PC, and we use it for gaming. Of course, people today use it for design, creative arts, and it does amazing AI work. The real breakthrough here is truly incredible: GeForce enabled AI— it let Geoff Hinton, Ilya Sutskever, and Alex Krizhevsky train AlexNet. We discovered AI, advanced AI, and then AI came back to GeForce to help with computer graphics. Here's the amazing part: out of 8 million pixels or so in a 4K display, we only compute 500,000 of them. The rest we use AI to predict— the AI guesses them. Yet the image is perfect. We inform the AI with the 500,000 pixels we computed (we ray-traced every single one, and they're beautiful, perfect), then we tell the AI: "If these are the 500,000 perfect pixels on this screen, what are the other 7.5 million?" And it fills in the rest perfectly. If you only have to compute fewer pixels, you can invest more in making those pixels higher quality. So the AI's extrapolation works exactly because you can focus your computing resources, attention, and energy on the 500,000 pixels. This is a perfect example of why AI is going to make us all superhuman: it does all the other work for us, allowing us to focus our time and energy on the really valuable things we do. We'll dedicate our own energy-intensive, attention-intensive resources to the few key pixels and use AI to upscale everything else. So this graphics card is now mostly powered by AI, and the computer graphics technology inside is incredible too.
This next one— as I mentioned earlier, in 2016 I built the first one for AI researchers, and we delivered the first one to OpenAI; Elon was there to receive it. This is a mini version, and the reason for that is because AI has now gone from AI researchers to everyone— to your PC or Mac. All of a sudden, you have your own AI supercomputer. Learning how to interact with AI is not unlike being someone who's really good at asking questions. If I want to be a better biologist, how do I use AI to do that? That question should be persistent, just like my generation grew up as the first to ask: "How can we use computers to do our jobs better?" The generation before us had no computers. I came into the industry before Windows 95— in 1984, there were no computers in offices. Shortly after that, computers started to emerge, so we had to ask ourselves how to use them to do our jobs better. The next generation won't have to ask that question, but they'll have to ask the obvious next one: "How can I use AI to do my job better?" Everybody really ought to just go try it. The crazy thing is, if I put a computer in front of someone who's never used one, they'll never learn it in a day— someone has to show them. But with ChatGPT, if you don't know how to use it, just type: "I don't know how to use ChatGPT, tell me." And it'll come back with examples. That's the amazing thing about intelligence: it will help you along the way and make you a superhuman in the process.
To think about this as a whole story: parallel processing reinvents modern gaming and revolutionizes an entire industry. Then that way of computing begins to be used across different industries. You invest in that by building CUDA, and then CUDA and the use of GPUs allows for a step change in neural networks and machine learning, beginning a revolution that's only increasing in importance today. All of a sudden, computer vision is solved, speech recognition is solved, language understanding is solved—these incredible problems associated with intelligence, one after another, problems we had no solutions for in the past, all of a sudden getting solved every couple of years. It's incredible. So you saw that in 2012, looked ahead, believed that's the future we're living in now, and made really big bets with high stakes to get there. My perception as a layperson is that it took a pretty long time—10 years. If AlexNet happened in 2012, and this audience is probably seeing and hearing so much more about AI and Nvidia specifically 10 years later, why did it take a decade? And also, because you had placed those bets, what did the middle of that decade feel like for you?
提问者（中文）：
把这一切串联起来看：并行处理重塑了现代游戏，彻底改变了整个行业。然后这种计算方式开始被应用到不同的行业。你们通过打造 CUDA 进行投资，而 CUDA 以及 GPU 的使用让神经网络和机器学习实现了跨越式发展，开启了一场如今重要性日益提升的革命。突然间，计算机视觉问题解决了，语音识别问题解决了，语言理解问题也解决了 —— 这些与智能相关的不可思议的问题，一个接一个地被攻克，那些我们过去束手无策的问题，如今每隔几年就会有一个被解决。这太不可思议了。所以你在 2012 年就看到了这一点，展望未来，相信那就是我们现在所处的时代，并为此下了风险极高的大赌注。作为门外汉，我感觉这花了相当长的时间 —— 整整十年。如果 AlexNet 是在 2012 年出现的，而现在十年过去了，观众们可能才更多地听到和看到关于人工智能以及英伟达的消息，为什么这需要十年时间？另外，因为你下了那些赌注，那十年中期的时候，你的感受是怎样的？
黄仁勋（英文）：
That's a good question. To me, there's always some problem, always a reason to be happy where you are, and always many reasons to carry on. I think as I was reflecting a second ago, that time sounds like just this morning. But in all things we pursue, first you have to have core beliefs. You have to reason from your best principles—ideally from physics, deep understanding of the industry, or science; reasoning from first principles. At some point, you have to believe something. If those principles and assumptions don't change, there's no reason to change your core beliefs. Along the way, there's always some evidence of success that you're heading in the right direction. Sometimes you go a long time without that evidence, might have to course-correct a little, but if you feel like you're on the right path, you keep going. The answer to why we stayed so committed is the opposite: there was no reason not to be committed because we believed it. I've believed in Nvidia for 30 plus years, and I'm still here working every single day. There's no fundamental reason to change my belief system—I fundamentally believe the work we're doing in revolutionizing computing is as true today, even more true, than it was before. So we'll stick with it until we have reason not to. Of course, there were very difficult times along the way—investing in something nobody else believes in, costing a lot of money, investors or others wanting you to just keep profits or improve the share price. But you have to believe in your future, invest in yourself. We believed so deeply that we spent tens of billions of dollars before it really took off. It was 10 long years, but it was fun along the way.
How would you summarize those core beliefs? What is it that you believe about the way computers should work and what they can do for us that keeps you going through that decade and doing what you're doing now, making bets for the next few decades?
The first core belief was our initial discussion about accelerated computing—parallel computing versus general-purpose computing. We believed combining those processors for accelerated computing was the way to go, and I still believe that today. The second was recognizing that deep learning networks (DNNs), which came to the public in 2012, have the ability to learn patterns and relationships from many types of data. They can learn more nuanced features if they're made larger—deeper or wider—and the scalability of the architecture is empirically true. The fact that larger model and data sizes lead to more knowledge is also empirical. So if that's the case, what are the limits? There's no physical, architectural, or mathematical limit that was ever found, so we believed we could scale it. The only other question is: what can you learn from data? Data is basically digital human experience. You can learn object recognition from images, speech from audio, even languages—vocabulary, syntax, grammar—just by studying letters and words. We've now demonstrated that deep learning can learn almost any modality of data and translate between modalities. What does that mean? Text to summary, text to text translation, text to image, image to captioning, amino acid sequences to protein structures. In the future, you'll go from protein to words—what does this protein do? Or give me a protein with these properties—identifying a drug target. You can see all these problems are around the corner to be solved. You can go from words to video. Why not words to action tokens for a robot? From the computer's perspective, how is that any different? It opened up a universe of opportunities and problems to solve, and that excites us.
It feels like we are on the cusp of this truly enormous change. When I think about the next 10 years, unlike the last 10, I don't think I can predict how I'll be using the technology currently being developed.
That's exactly right. I think the last 10 years were really about the science of AI. The next 10 years will have plenty of AI advancements, but it's going to be the application science of AI—fundamental science versus applied science. The application side becomes: how do I apply AI to digital biology? To climate technology? To optimizing logistics? To healthcare? To education?
I'd love to choose a couple of those to help people see how this fundamental change in computing is actually going to change their daily lives, how they'll use technology based on everything we just talked about. One thing I've heard you talk a lot about, and I have a particular interest in, is physical AI—robots, meaning humanoid robots, but also self-driving cars, smart buildings, autonomous warehouses, autonomous lawnmowers, and more. From what I understand, we might be about to see a huge leap in what these robots are capable of because we're changing how we train them. Up until recently, you had to train robots in the real world where they could get damaged or wear down, or get data from limited sources like humans in motion capture suits. That means robots aren't getting as many examples as they need to learn quickly. But now we're training robots in digital worlds, meaning way more repetitions a day, more conditions, learning faster. So we could be in a big bang moment for robots right now, and Nvidia is building tools to make that happen. You have Omniverse, which I understand is 3D worlds that help train robotic systems so they don't need to train in the physical world. That's exactly right. You just announced Cosmos, which makes that 3D universe much more realistic, so you can get all kinds of scenarios—different lighting, times of day, experiences for the robot to go through—so it can get even more out of Omniverse. As a kid who grew up loving Isaac Asimov's stories and dreaming about a future with robots, how do we get from the robots we have now to the future world of robotics you see?
提问者（中文）：
我想选几个领域，帮助人们了解计算领域的这种根本性变革实际上将如何改变他们的日常生活，以及他们将如何使用基于我们刚才所谈论的一切的技术。有一件事我经常听你提到，而且我特别感兴趣，那就是物理人工智能 —— 机器人，既包括人形机器人，也包括自动驾驶汽车、智能建筑、自动仓库、自动割草机等等。据我了解，我们可能即将看到这些机器人的能力实现巨大飞跃，因为我们正在改变训练它们的方式。直到最近，你还必须在现实世界中训练机器人，它们可能会受到损坏或磨损，或者只能从有限的来源获取数据，比如穿着动作捕捉服的人类。这意味着机器人无法获得足够的示例来快速学习。但现在我们在数字世界中训练机器人，这意味着每天可以进行更多的重复训练，更多的场景模拟，学习速度更快。所以我们现在可能正处于机器人领域的大爆炸时刻，而英伟达正在打造工具来实现这一点。你们有 Omniverse，据我了解，这是一个 3D 世界，可以帮助训练机器人系统，这样它们就不需要在物理世界中进行训练。完全正确。你们刚刚宣布了 Cosmos，它让那个 3D 宇宙变得更加逼真，这样就可以获得各种各样的场景 —— 不同的光线、一天中的不同时间、机器人需要经历的各种情况 —— 这样它就能从 Omniverse 中获得更多东西。作为一个从小就喜欢艾萨克・阿西莫夫的故事并梦想着机器人未来的孩子，我们如何才能从现在的机器人发展到你所看到的未来机器人世界？
黄仁勋（英文）：
Let me use ChatGPT as a reference to explain Omniverse and Cosmos. First, when ChatGPT came out, it was extraordinary—generating text from prompts. But as amazing as it was, it had a tendency to hallucinate if it went on too long or pontificated about an uninformed topic, generating plausible but untrue answers. It wasn't grounded in truth. The next generation shortly after could be conditioned by context—you could upload a PDF, and the PDF became ground truth; it could look up search results, and that became ground truth. Between that, it could reason about how to produce the answer you're asking for. So first, generative AI; second, grounding it in truth. Now, moving to the physical world: we need a foundation model, just like ChatGPT's core foundation model—that was the breakthrough. For robotics to be smart about the physical world, it has to understand gravity, friction, inertia, geometric and spatial awareness, object permanence (that an object is still there when you look away and come back), cause and effect (if you tip something, it falls over). These physical common sense things have to be captured in a world foundation model so AI has world common sense. That's what we did with Cosmos—created a world model, just like ChatGPT is a language model. The second thing: we need to ground Cosmos in truth, just like we did with PDFs and ChatGPT. The way we augment Cosmos with ground truth is through physical simulations. Omniverse uses physics simulations based on principled solvers—Newtonian physics, the math we know. All fundamental physics laws are encoded in Omniverse, which is why it's a simulator. Using the simulator to ground Cosmos, we can now generate an infinite number of physically plausible future scenarios, and robots learn from them before entering the physical world. To illustrate: if you have a factory robot and want it to learn every route, instead of manually testing all routes (taking days, wearing down the robot), we can simulate all of them digitally in a fraction of the time, in many situations—dark, blocked, etc. The robot learns much faster.
Cosmos这是英伟达 2025 年在 CES 展开源的生成式世界基础模型平台，核心是为机器人、自动驾驶等物理 AI 开发提供逼真的合成数据与虚拟训练环境。它基于 Transformer 的自回归和扩散双架构，能用文本、图像等多模态输入生成模拟物理规律的视频，还能预测未来世界状态。其训练用到 9000 万亿 token 的数据，涵盖 2000 万小时自动驾驶、机器人等领域数据，能生成暴雨、暴雪等难在现实中复刻的场景。它有 Nano、Super、Ultra 三个型号适配不同部署需求，像小鹏汽车、Uber 等企业已用它加速机器人和自动驾驶开发，搭配 Omniverse 时还能成倍生成合成训练数据，大幅缩短物理 AI 的开发周期。
Omniverse这是英伟达打造的用于构建数字孪生与物理 AI 仿真的平台，可理解为机器人等智能设备的 “虚拟训练场”。它依托 OpenUSD 技术和 RTX 渲染能力，能创建高度还原物理规律的 3D 虚拟场景（即数字孪生）。在机器人训练方面，开发者可在上面搭建工厂、仓库等场景，模拟机器人的搬运、分拣等动作并测试优化；在医疗领域，它能实现心脏血管系统的 3D 可视化，还可仿真医疗设施以优化医院运营规划；工业领域中，企业能借助它模拟多机器人集群作业，提前排查实际部署可能出现的问题。该平台还推出了托管服务和预构建工作流，方便开发者快速开发物理 AI 相关应用。
If you play this out 10 years, how do you see people actually interacting with this technology in the near future?
黄仁勋：
Everything that moves will be robotic someday, and it will be soon. The idea of pushing a lawnmower is already kind of silly—maybe people do it because it's fun, but there's no need to. Every car is going to be robotic, humanoid robots too. The technology necessary to make it possible is just around the corner. So everything that moves will be robotic, and they'll learn how to be a robot in Omniverse and Cosmos. We'll generate all these plausible, physically plausible futures, and the robots will learn from them before coming into the physical world. A future where you're just surrounded by robots is certain. And I'm just excited about having my own R2-D2. Of course, R2-D2 wouldn't be quite the can that rolls around—it'll probably be a different physical embodiment, but the idea is the same. So my R2 is going to go around with me—sometimes it's in my smart home, sometimes in my office, sometimes in my car. R2 is with me all the time. When I get home, where I left a physical version of R2, I'll interact with it. I think the idea that we'll have our own R2-D2 for our entire life, growing up with us, is a certainty now.
提问者：
I think a lot of news coverage when they talk about futures like this focuses on what could go wrong, and that makes sense—there's a lot that could go wrong. We should talk about what could go wrong so we can keep it from happening. That's the approach we like to take on the show: what are the big challenges so we can overcome them? What buckets do you think about when you're worrying about this future?
黄仁勋：
Well, there's a whole bunch of stuff that everybody talks about—bias, toxicity, hallucination (speaking with great confidence about something it knows nothing about, and we rely on that information). That's a version of generating fake information—fake news, fake images, whatever. Of course, impersonation: AI does such a good job pretending to be a human, it could do an incredibly good job pretending to be a specific human. The spectrum of areas we have to be concerned about is fairly clear, and a lot of people are working on it. Some of the stuff related to AI safety requires deep research and engineering—cases where AI wants to do the right thing but just didn't perform correctly, resulting in harm. For example, a self-driving car that wants to drive properly but has a sensor breakdown, fails to detect something, or makes an overly aggressive turn. That's a whole bunch of engineering that has to be done to ensure AI safety by making sure the product functions properly. Lastly, what if the system fails even when AI wants to do a good job? Like AI wants to stop something from happening, but the machine breaks down just when it needs to act. This is no different than a flight computer in a plane having three versions—triple redundancy in autopilots, plus two pilots, air traffic control, and other pilots watching out. So AI safety systems have to be architected as a community such that these AIs first function properly; when they don't, they don't put people in harm's way; and there are sufficient safety and security systems around them to keep AI safe. The conversation around this is gigantic, and we have to break it down and build solutions as engineers.
提问者：
One of the incredible things about this moment we're in right now is that we no longer have a lot of the technological limits that we had in a world of CPUs and sequential processing. We've unlocked not only a new way to do computing but also a way to continue to improve—parallel processing has a different kind of physics to it than the improvements we were able to make on CPUs. I'm curious about the scientific or technological limitations that we face now in the current world that you're thinking a lot about.
黄仁勋：
Everything in the end is about how much work you can get done within the limitations of the energy you have. That's a physical limit—the laws of physics around transporting information, flipping bits, and moving bits. At the end of the day, the energy it takes to do that limits what we can accomplish, and the amount of energy we have limits what we can do. But we're far from having any fundamental limits that keep us from advancing. In the future, we seek to build better and more energy-efficient computers. This little device here— the big version was $250,000? This is a baby AI supercomputer. It lets you work far from home and work, live further away. You ask yourself questions like: What would happen if I had a software programmer with me all the time, who could write any software I dream up? Would that change my opportunities? What would it free me to do? I think in the next decade, intelligence—not for everything, but for some things—will basically become superhuman. I can tell you exactly what that feels like: I'm surrounded by superhuman people—superintelligence from my perspective—because they're the best in the world at what they do, and they do it way better than I can. I'm surrounded by thousands of them. So suppose now everybody is surrounded by these super AIs that are very good at specific things. How would that make you feel? It's going to empower you, make you feel confident. I'm pretty sure you use ChatGPT or other AIs, and I feel more empowered, more confident to learn something new today. The barriers to accessing almost any particular knowledge have been reduced. I have a personal tutor with me all the time. I think that feeling should be universal, and if there's one thing I'd encourage everybody to do, it's to get yourself an AI tutor right away. That AI tutor can teach you anything—help you program, write, analyze, think, reason. It's going to really make you feel empowered. I think that's our future: we're going to become superhumans not because we have superpowers, but because we have super AIs.
提问者：
Could you tell us a little bit about each of these objects?
黄仁勋：
This is a new GeForce graphics card—the RTX 50 series. It's essentially a supercomputer that you put into your PC, and we use it for gaming. Of course, people today use it for design, creative arts, and it does amazing AI work. The real breakthrough here is truly incredible: GeForce enabled AI— it let Geoff Hinton, Ilya Sutskever, and Alex Krizhevsky train AlexNet. We discovered AI, advanced AI, and then AI came back to GeForce to help with computer graphics. Here's the amazing part: out of 8 million pixels or so in a 4K display, we only compute 500,000 of them. The rest we use AI to predict— the AI guesses them. Yet the image is perfect. We inform the AI with the 500,000 pixels we computed (we ray-traced every single one, and they're beautiful, perfect), then we tell the AI: "If these are the 500,000 perfect pixels on this screen, what are the other 7.5 million?" And it fills in the rest perfectly. If you only have to compute fewer pixels, you can invest more in making those pixels higher quality. So the AI's extrapolation works exactly because you can focus your computing resources, attention, and energy on the 500,000 pixels. This is a perfect example of why AI is going to make us all superhuman: it does all the other work for us, allowing us to focus our time and energy on the really valuable things we do. We'll dedicate our own energy-intensive, attention-intensive resources to the few key pixels and use AI to upscale everything else. So this graphics card is now mostly powered by AI, and the computer graphics technology inside is incredible too.
This next one— as I mentioned earlier, in 2016 I built the first one for AI researchers, and we delivered the first one to OpenAI; Elon was there to receive it. This is a mini version, and the reason for that is because AI has now gone from AI researchers to everyone— to your PC or Mac. All of a sudden, you have your own AI supercomputer. Learning how to interact with AI is not unlike being someone who's really good at asking questions. If I want to be a better biologist, how do I use AI to do that? That question should be persistent, just like my generation grew up as the first to ask: "How can we use computers to do our jobs better?" The generation before us had no computers. I came into the industry before Windows 95— in 1984, there were no computers in offices. Shortly after that, computers started to emerge, so we had to ask ourselves how to use them to do our jobs better. The next generation won't have to ask that question, but they'll have to ask the obvious next one: "How can I use AI to do my job better?" Everybody really ought to just go try it. The crazy thing is, if I put a computer in front of someone who's never used one, they'll never learn it in a day— someone has to show them. But with ChatGPT, if you don't know how to use it, just type: "I don't know how to use ChatGPT, tell me." And it'll come back with examples. That's the amazing thing about intelligence: it will help you along the way and make you a superhuman in the process.
提问者：
All right, I have one more question if you have a second. This is not something I planned to ask, but on the way here, I'm a little bit afraid of planes— it's not my most reasonable quality— and the flight here was a little bumpy, very bumpy. I was sitting there, and the plane was moving, and I was thinking about what they'd say at my funeral. After it was over, I thought, "If that's what the tombstone says— 'She asked you questions'— that's fine." I loved my husband, friends, and family, but the thing I hoped they'd talk about was optimism. I hope they'd recognize what I'm trying to do here. And I'm very curious for you: you've been doing this a long time, and there's so much you've described in this vision ahead. What would the theme be that you'd want people to say about what you're trying to do?
黄仁勋：
Very simply: "They made an extraordinary impact." I think we're fortunate because of some core beliefs we had a long time ago, sticking with those beliefs, and building upon them— we found ourselves today being one of the most important and consequential technology companies in the world, potentially ever. We take that responsibility very seriously, and we work hard to make sure the capabilities we've created are available to large companies as well as individual researchers and developers, across every field of science— no matter if it's profitable or not, big or small, famous or otherwise. It's because of this understanding of the consequential work we're doing and the potential impact it has on so many people that we want to make this capability as pervasive as possible. I do think that when we look back in a few years, and I hope that what the next generation realizes— well, first of all, they'll know us because of all the gaming technology we create— but I do think they'll look back, and in hindsight, realize there was this company almost at the epicenter of all these changes, like remote work and the AI revolution.
黄仁勋访谈
Thank you so much for your I'm so happy to do it.Before we dive I wanted to tell you how this interview is going to be a little bit different than other interviews I've seen you doing okay? I'm not going to ask you any questions about you could ask company thank you. I'm not going to ask you questions about your management style or why you don't like one on ones. I'm not going to ask you about regulations or I think all of those things are but I think that our audience can get them well covered okay? What we do on huge if true is we make optimistic explainer videos and we've covereded I'm the worst person to be an explainer video. I think you might be the best and I and that's what I'm really hoping that we can do is make a joint explainer video about how can we actually use technology to make the future and we do it because we believe that when people see those better they help build them. So the people that you're going to be talking to are they are optimists who want to build those better but but because we cover so many different.A.Year old video on the Nvidia YouTube channel where the mythbusters they use a little robot shooting paintballs one by one to show solving problems one at a time or sequential processing on a CPU but then they roll out this huge robot that shoots all of the paintballs at once doing smaller problems all at the same time or parallel processing on a GPU 321.So Nvidia unlocks all of this new power for video games gaming. First the video games requires parallel processing for processing 3D graphics and we chose video games because one we loved the application it's a simulation of virtual worlds and who doesn't want to go to virtual worlds and and we had the good observation video games has potential to be the largest market for entertainment ever and it turned out to be true and having it being a large market is important because the technology is complicated.And if we had a large market our r0 d budget could be large we could create new technology.And that flywheel between technology and market and greater technology was was really the flywheel that got Nvidia to become one of the most important technology companies in the world. It was all because of video games.I've heard you say that gpus a time machine? Yeah? Could you tell me more about what you meant by that? A GPU is like a time machine because it lets you see the future sooner. One of the most amazing things anybody's ever said to me was a quantum chemistry scientist he said Jensen. Because of Nvidia's I can do my life's work.In my lifetime that's time travel. He was able to do something that was beyond his lifetime within his lifetime and and this because we make applications run so much faster.And so you get to see the future. And so when you're doing weather prediction for example you're seeing the future when you're doing a simulation a virtual city with virtual traffic and we're simulating our self driving car through that virtual city we're doing time travel. So parallel processing takes off in gaming and it's allowing us to create worlds in computers that we never could have before and and gaming is sort of this this first incredible.Case of parallel processing unlocking a lot more power and then as you said people begin to use that power across many different industries. The case of the of the quantum chemistry researcher.When I've heard you tell that story it's that he was running molecular simulations in a way where it was much faster to run in parallel on Nvidia gpus even then than it was to run them on the supercomputer with the CPU that he had been using before. That's true. So oh my god it's revolution revolutionizing all of these other industries as well. It's beginning to change how we see what's possible with computers and.My understanding is that in the early 2000 you see this and you realize that actually doing that is a little bit because what that researcher had to do was he had to sort of trick the gpus into thinking that his problem was a graphics problem. That's exactly right. that's very good. so you you did some so you create a way to make that a lot easier? That's right. it's a platform called which lets programmers tell the GPU what to do using programming languages that they already like c0. That's a big deal because it gives way more people easier access to all of this computing power. Could you explain what the vision was that led you to create cuda?Partly researchers discovering it partly internal inspiration and.I am partly solving a problem and you know a lot of interesting interesting ideas come out of that soup.You know some of it is is aspiration and inspiration some of it is just desperation you know and and so in in the case of kudos very much this the same way and probably the first external ideas of using our gpus for parallel processing emerged out of some interesting work in medical imaging.A couple of researchers at mass general were using it to do CT reconstruction. They were using our graphics processors for that reason and it inspired us. Meanwhile the problem that we're trying to solve inside our company.Has to do with the fact that when you're trying to create these virtual worlds for video games you would like it to be beautiful but also dynamic.Water should.Flow like water and explosions should be like explosions. So there's particle physics you want to do fluid dynamics you want to do and and that is much harder to do if your pipeline is only able to do computer graphics. So we have a natural reason to want to do it in in the the market that we were serving researchers were also.Um horsing around with using our gpus for a general purpose acceleration and so there were there are multiple multiple factors that were coming together in that soup. We just when the time came and and we decided to to do something proper and create a kudos a result of that fundamentally the reason why why I was certain that cuda was going to be successful and we we put the whole company behind it was because.Fundamentally.Our GPU was going to be the highest volume parallel processors built in the world because the market of video games was so large and so this architecture has a good chance of reaching many people. It has seemed to me like creating kudo was this incredibly optimistic.Huge if true thing to do where you were saying.If we create a way for many more people to use much more computing they might create incredible things and then of course it came true. They did! In 12 a group of three researchers submits an entry to a famous where the goal is to create computer systems that could recognize images and them with categories and their entry just crushes the competition. It gets way fewer answers. Wrong. It was incredible. It blows everyone away. It's called alexnet and it's a kind of AI called a neural network. My understanding is one reason it was so good is that they used a huge amount of data to train that system and they did it on Nvidia gpus. All of a sudden gpus weren't just a way to make computers faster and more they're becoming the engines of a whole new way of computing. We're moving from instructing computers with step by step directions to training computers to learn by showing them a huge number of examples. This moment in 2012 really kicked off this truly seismic shift that we're all seeing.With AI right could you describe what that moment was like from your perspective and what did you see it would mean?For all of our futures when you create something new like cuda if you build it they might not come.And that's that's always the the cynics perspective however the optimist perspective would say but if you don't build it they can't come.And that's usually how we look at the world you know we we have to reason about intuitively why this would be very useful and in fact in 2012 AIA sus cover and Alex krusywski and geoff hinton in the university of Toronto the lab that they were at they reached out to AG force gtx 580 because they learned about cuda and that cuda might be able to be used as a parallel processor for training alexnet and so our inspiration that g force could be the the vehicle to bring out this.Parallel architecture into the world and that researchers would somehow find it someday was a good was a good strategy. It was a strategy based on hope but it was also reasoned hope. The thing that really caught our attention was simultaneously we were trying to solve the computer vision problem inside the company and we we were trying to get cuda to be a good computer vision processor and we were frustrated by by a whole bunch of early developments internally with respect to our our computer vision effort and getting cuda to be able to do it and all of a sudden we saw alexnet.Um this new algorithm that that is completely different than computer vision algorithms before it take a giant leap in terms of capability for computer vision and when we saw that it was partly out out of interest but partly because we were struggling with something ourselves and so we were we were highly interested to want to see it work and so when we when we looked at alexnet we were inspired by that but the big breakthrough I would say.Ah is when we when we saw alexnet we asked ourselves you know how far can alexnet go.If it can do this with computer how far can it go?And if it if it could go to the limits of what we think it could go the type of problems it could solve what would it mean for the computer industry and what would it mean for the computer architecture.And we were we were um ah we rightfully reasoned that if machine learning of these deep learning architecture can scale the vast majority of machine learning problems could be represented with deep neural networks and the type of problems we could solve with machine learning is so vast that it has the potential of reshaping the computer industry altogether.And which prompted us to reengineer the entire computing stack which is where DG X came from and this little baby DGX sitting here all of this came from from that observation that we ought to reinvent the entire computing stack layer by layer by layer. You know computers after 65 years since IBM system 360 introduced modern general purpose computing we've reinvented computing as we know it.To think about this as a whole story so parallel processing reinvents modern gaming and revolutionizes an entire industry. Then that way of computing that parallel processing begins to be used across different industries. You invest in that by building cuda and then.Kuda and the use of gpus.Allows for a ah step change in neural networks and machine learning and begins a sort of revolution that we're now seeing.Ah only increase in importance today.All of a sudden computer vision is solved all of a sudden speech recognition solved all of a sudden language understanding is solved these incredible problems associated with intelligence one by one by one by one.Where we had no solutions for in the past desperate desire to have solutions for all of a sudden one after another gets solved you know every couple of years. It's incredible. Yeah so you're seeing that in 2012 you're looking ahead and and believing that that's the future that you're going to be living in now.And you're making bets that get you there really big bets that have very high stakes. And then my perception as a layperson is that it takes a pretty long time to get there. You make these years ah 10 years.So my question is.If Alex that happened in twenty twelve and this audience is probably seeing and hearing so much more about AI and Nvidia specifically 10 years later.Why did it take a decade? And also because you had placed those what did the middle of that decade feel like for you? Well that's a good question. It probably felt like you to me to me there's always some problem and then there's some reason to to be there's always some reason to be.Ah happy about where you are and and there's always many reasons to carry on and so so I think as I was reflecting a second ago that sounds like this morning and so but I would say that in all things that we pursue first you have to have core beliefs.You have to reason from from your best principles um and ideally you're reasoning from it from principles of either physics or deep understanding of of of the industry or deep understanding of the science wherever you're reasoning from you reason from first principles.And at some point you have to believe something. If those principles don't change and the assumptions don't change then you there's no reason to change your core beliefs and then along the way there's always some.Ah evidence of of um you know of success and and that you're you're leading in the right direction and sometimes you know you go a long time without evidence evidence of success and you might have to course correct a little but the evidence comes in and if you feel like you're going the right direction we just keep on going. The question of why did we stay so committed for so long the answer is actually the opposite.There was no reason to not be committed because we are we believed it.And and um I've believed in Nvidia for 30 plus years and and I'm still here working every single day and there's no fundamental reason for me to change my belief system and I fundamentally believe that the the work we're doing in revolutionizing computing is as true today even more true today than it was before 000 so we we'll stick with it you know until until otherwise.Um there's of course.Very difficult times along the way you know when you're investing in something and nobody else believes in it and costs a lot of money and you know maybe investors or or others would rather you just keep the profit or you know whatever it is improve the share price or whatever it is but you have to believe in your future you have to invest in yourself and and.We believe so deeply that that we we you tens of billions of dollars before before it really happened I it was it was 10 long but it was it was it was fun along the way. How would you summarize those core beliefs? What is it that you believe about the way computers.Should work and what they can do for us that keeps you not only coming through that decade but also doing what you're doing now making bets I'm sure you're making for the next few decades.The first core belief was our first discussion was about accelerated computing parallel computing versus versus general purpose computing. We would add two of those processors together and we would do accelerated computing and I continue to believe that today. The second was was the recognition that these deep learning networks these dnns that came to the public during 2012 these deep neural networks have the ability to learn patterns and relationships from a whole bunch of different types of data and that it can learn more and more nuanced features if it could be larger and larger and it's easier to make them larger and larger to make them deeper and deeper or wider and wider and so the scalability of the architecture is is empirically true.the fact that model size and the data size being larger and you can learn more knowledge is also empirically and so if that's the you you what are the limits? There's unless there's a physical limit or an architectural limit or a mathematical limit and it was never found and so we believed that you could scale it. Then the question the only other question is what can you learn from data.What can you learn from experience? Data is basically digital versions of human experience and so what can you learn? You obviously can learn object recognition from you can learn speech from just listening to you can learn even languages and vocabulary and syntax and grammar and all just by studying a whole bunch of letters and words. So we've now demonstrated that or deep has the ability to learn almost any modality of data.And it can translate to any modality of and so what does that mean? You can go from text to right? Summarize a paragraph. You can go from text to text. Translate from language to language. You can go from text to that's image you can go from that's captioning.You can even go from amino acid sequences to protein structures in the future you'll go from protein to words. What does this protein do.Or give me an example of a protein that has these properties.You know identifying a drug target and so you could just see that all of these problems are around the corner to be solved.Ah you can go from words.To video.Why can't you go from words to action tokens for a robot?You from the computer's how is it any different? And so it opened up this universe of opportunities and universe of problems that we can go solve and that gets us quite excited.It feels like we are on the cusp of this.Truly enormous change when I think about the next 10 years I.Unlike the last 10 years I know we've gone through a lot of change already but I don't think I can predict anymore how I will be using the technology that is currently being developed. That's exactly right. I think the last tent the reason why you feel that way is the last 10 years was really about the science of AI.The next 10 years we're going to have plenty of signs of AI but the next 10 years is going to be the application science of AI.The fundamental science versus the application science and so the applied the application side of AI now becomes how can I apply AI to digital biology? How can I apply AI to climate technology? How can I apply AI to the to to optimizing logistics? How can I apply AI you how DOI apply AI you right? And so.I'd love to choose a couple of those to help people see how this fundamental change in computing that we've been talking about is actually going to change their experience of their lives how they're actually going to use technology that is based on everything we just talked about. One of the things that I've now heard you talk a lot about and I have a particular interest in is physical AI or in other words robots my friends meaning humanoid robots but also robots like self driving cars and smart buildings or autonomous warehouses or autonomous lawn moowers or more. From what I understand we might be about to see a huge leap in what all of these robots are capable of because we're changing how we train them. Up until recently you've either had to train your robot in the real world where it could get damaged or wear down or you could get data from fairly limited sources like humans in motion capture suits.That means that robots aren't getting as many examples as they'd need to learn more quickly but now we're starting to train robots in digital worlds which means way more repetitions a day way more conditions learning way faster so we could be in a bigbang moment for robots right now and Nvidia is building tools to make that happen.You have omniverse and my understanding is this is 3D worlds that help train.Robotic systems so that they don't need to train in the physical world? That's exactly right. You just announced which is ways to make that 3D universe much more realistic so you can get all kinds of if we're training something on this many different kinds of lighting on the many different times of many you experiences for the robot to go through so that it can get even more out of omniverse.As a kid who grew up loving data on star Isaac asimov's and just dreaming about a future with how do we get from the robots that we have now to the future world that you see of robotics? let me language maybe chatgpt as a reference for understanding omniverse and cosmos.And so so first of all when chatgpt first came out it was it was extraordinary and it has the ability to do to basically from your prompt generate text however as amazing as it was.It has the tendency to hallucinate if it goes on too long or if it pontificates about a topic you know is not informed about it'll still do a good job generating plausible answers. It just wasn't grounded into truth and so.So um ah people people call it hallucination.And so the the next generation ah shortly it was had the ability to be conditioned by context. So you could upload your PDF and now it's grounded by the PDF. The the PDF becomes the ground truth it could be it could actually look up search and then the search becomes its ground truth and between that it could reason about what is how to produce the answer that you're asking for.And so the first part is a generative AI and the second part is ground okay? And so now let's come into the physical the world model. We need a foundation model just like we need chatgpt had a core foundation model. That was the breakthrough. In order for robotics to to be smart about the physical it has to understand things like gravity.Friction inertia geometric and spatial awareness it has to.Understand that an object is sitting there even when I looked away when I come back it's still sitting there. Object permanence it has to understand cause and effect. If I tip it it will fall over and so so these kind of physical common sense if you will has to be.Captured or encoded into a world foundation model so that the AI has world common sense.Okay? And so we have to somebody has to go create that and that's what we did with we created world language model.Just like chatgpt was a language model this is a world model. The second thing we have to go do is we have to do the same thing that we did with pdfs and context and grounding it with ground truth and so the way we augment cosmos with ground truth is with physical simulations.Because omniverse uses physics simulation which is based on principled solvers. The mathematics is newtonian physics is the right it's the math we know.Okay all of the the fundamental laws of physics we've understood for a very long time and it's encoded into captured into omniverse. That's why omniverse is a simulator.And using the simulator to.Ground or to cosmos can now generate an infinite.Number of stories of the future and they're grounded on physical truth.Just like between PDF or search plus chatgpt we can generate an infinite.Amount of interesting things answer a whole bunch of interesting questions. The combination of omniverse plus cosmos you could do that the physical world. So to illustrate this for the audience if you had a robot in a factory and you wanted to make it learn every route that it could take instead of manually going through all of those routes which could take days and could be a lot of wear and tear on the robot we're now able to simulate all of them digitally in a fraction of the time and in many different situations that the robot might face. It's dark it's blocked it's etctera. The robot is now learning much much faster. It seems to me like the future might look very different than today.If you play this out 10 years how do you see people actually interacting with this technology in the near future.Clio everything that moves will be robotic someday and it will be soon. You know the the idea that'll be pushing around a lawnmower is already kind of silly.You know maybe people do it because because it's fun but but there's no need to and and um every car is going to be robotic humano robots.The technology necessary to make it possible is just around the corner and.So everything that moves will be robotic and they'll they'll learn how to be a robot in omniverse cosmos and will generate.All these plausible physically plausible futures and the robots will learn from them.And then they'll come into the physical world and you know it was exactly the same.future where? you're just surrounded by robots is for certain. And I'm just excited about having my own r2 d 2. Of course r2 d2 wouldn't be quite the can that it is and roll roll around. It'll be you know r2 it'll probably be a different physical embodiment.but it's always you know? So my r2 is going to go around with sometimes it's in my smart sometimes it's in my sometimes it's in it's in my car. So r2 is with me all the you when I get you where I left? a physical version of r 2 you whatever whatever that version happens to you know will interact with our 2 so I think the idea that we'll have our own d2 for our entire life.And it grows up with us's a certainty now? I think a lot of.News when they talk about futures like they focus on what what could go wrong and that makes sense. there so is a lot that could go wrong? we should talk about what could go wrong so we can keep it from from going wrong that's the approach that we like to take on the show is what are the big challenges? so that we can overcome them. what? Buckets do you think about when you're worrying about this future.Well there's there's a whole bunch of the stuff that everybody talks about bias or toxicity or or just hallucination you know speaking with great confidence about something it knows nothing about and as a result we rely on that information.Um ah generating that's a version of generating fake information fake fake news or fake images or whatever it is. Of course impersonation. It does such a good job pretending to be a human. It could be it could do an incredibly good job pretending to be a specific human. And so so the the um the the spectrum of of areas that we have to be concerned about is fairly clear and there's a lot of there's a lot of people who are working on it. There's there's a.Some of the stuff some of the stuff related to AI safety requires deep research and deep deep engineering and that's simply it wants to do the right thing it just didn't perform it right and as a result hurt somebody. You know for example a self driving car that wants to drive nicely and and drive properly and just somehow the sensor broke down or or it didn't detect something or you know made it to too aggressive turn or whatever it is it did it poorly it did it wrongly wrongly and and so that's that's a whole bunch of engineering that has to be done to to make sure that AI safety is upheld by making sure that the product functioned properly and then and then lastly you know whatever what happens if the system that the AI wants to do a good job the system failed meaning the AI wanted to stop um stop stop something from happening and it turned out just when it wanted to do it the machine broke down.And so this is no different than than a flight computer inside a plane having three versions of them and then so there's there's triple redundancy inside the system inside autopilots and then you have two pilots and then you have air traffic control and then you have other pilots watching out for these pilots and so so that the AI safety systems has to be architected as a community such that.Such that these AIS one work work function function properly. When they don't function properly they don't put people in harm's way and and that they's sufficiently safety and security systems all around them to make sure that that we keep AI safe. And so there's the spectrum of conversation is gigantic and and um you know we have to take the parts take the parts apart and and build them as engineers.One of the incredible things about this moment that we're in right now is that we no longer have a lot of the.Technological limits that we had in a world of cpus and.In sequential processing and we have.Unlocked not only and a new way to do computing and.But also a way to continue to improve parallel processing has AA different kind of physics to it than the improvements that we were able to make on cpus. I'm curious.Scientific or technological limitations that we face now in the current world that you're thinking a lot about.Everything in the end is about how much work you can get done within the limitations of the energy that you have.And so that that's ah that's a physical limit and the laws of physics about transporting sick information and.Transporting bits flipping bits and transporting bits. At the end of the day the energy it takes to do that um limits what we can get done and the amount of energy that we have limits what we can get done.We're far from having any fundamental limits that keep us from advancing. In the we seek to build better and more energy efficient computers. this little the big version of it was 250000 dollars? that's a little baby.Baby digits? this is an AI supercomputer.Okay to work far further away from home and from from ah from work and live further away and so so.You ask yourself kind of these these questions you know what would happen if if um.I have a software programmer with me all the time and whatever it is I I can dream up the software programmer could write for me.Would that change my opportunity? you know what? What is a free me to be able to do and and so on and so forth and so so I think the the next the next decade intelligence not for but for for some things would basically become superhuman and and so.So um I.I can tell you exactly what that feels like.I'm surrounded by superhuman people. A superintelligence from my perspective because they're the best in the world at what they do.And they do what they do way better than I can do it.And.And I'm surrounded by thousands of them.Things.And so so suppose suppose now everybody.Is surrounded by these super AIS that are very good at specific things or good at some of the things.What would that make you feel? Well it's going to empower it's going to make you feel confident and I'm pretty sure you probably use chatgpt and AI and I I feel more empowered more confident to learn something today. The knowledge of almost any particular the barriers to that it has been reduced.I have a personal tutor with me all of the time and and so I I think that that feeling should be universal and and I if if there's one thing that I would encourage everybody to do is to go get yourself an AI tutor right away. That AI tutor could of course just teach you things. Anything you like. Help you program help you write help you analyze help you think help you reason you know all of those things is gonna really make you feel empowered and and um.I think that that's that's going to be our future we're going to become.We're going to become superhumans not because we have super we're going to become superhumans because we have super AIS.Could you tell us a little bit about each of these objects.This is a new gforce graphics card and yep just yes.And this is the RTX 50 series. It is essentially a supercomputer that you put into your PC and we use it for gaming. Of course people today use it for design and creative arts and it does.Amazing AI and the the real breakthrough here and this is this is truly an amazing an amazing thing. G forcece enabled AI and enabled jeoff hinton iiaia suskever and Alex kruszewski to be able to train alexnet. We discovered AI and and we advanced AI then AI came back to g force.To help computer graphics and so here's the amazing thing.Out of 8 million pixels or so in a4 k display we are computing we're processing only 500000 of them. The rest of them we use AI to predict the AI guessed it.And yet the image is perfect we inform it by the five hundred zero pixels that we computed and we ray traced every single 1 it's all beautiful it's perfect and then we tell the AI if these are the five hundred zero perfect pixels in this screen what are the other 8 million.And it goes fills in the rest of the screen and it's perfect and if you only have to do fewer pixels are you able to invest more in doing that because.You have fewer to do so then the quality is better. So the extrapolation that the AI does exactly because whatever computing whatever attention you have whatever resources you have you can place it into 500000 pixels.Now this is a perfect example of why.AI is going to make us all superhuman because all of the other things that it can do it'll do for us allows us to take our time and energy and focus it on the the really really valuable things that we do.And so we'll take our own resource which is you know.Energy intensive attention intensive and well dedicated to the few 0 pixels and use AI to super resit up resit you know to everything else. So this this graphics card is now powered mostly by AI.And the computer graphics technology inside is incredible as well. And then this next one as I mentioned earlier in 2016 I built the first one for AI researchers and we delivered the first one to OpenAI and Elon was there to receive it and this this version I built a mini mini version and the reason for that is because AI has now gone from AI researchers to every.To your PC or a Mac and all of a sudden you have your own AI supercomputer.And and um I these learning how to interact with with AI is not unlike being someone who is really good at asking questions.If I want to be a I how DOI use AI to be a better biologist? That question should be persistent across just as my generation.Grew up as a first generation that has to ask ourselves how can we use computers to do our jobs better.Yeah the generation before us had no computers. My generation was the first generation that had to ask the question how DOI use computers to do my job better.Remember I came into the industry before Windows 95.1984 there were no computers in offices.And after that shortly after that computers started to emerge and so we had to ask ourselves how do we use computers to do our jobs better the next generation.Doesn't have to ask that question but it has to ask the obviously next question how can I use AI to do my job better.Everybody really ought to just go try you know? The thing that's really crazy is if I put a computer in front of somebody and they've never used a there is no chance they're going to learn that computer in a day.There's just no chance somebody really has to show it to you and yet with chatgpt if you don't know how to use it all you have to do is type in I don't know how to use chatgpt tell me.And it would come back and give you some examples.And so that's the amazing thing you know this.The the amazing thing about intelligence is it will help you along the way and make you a superhuman you know along the way. All right I have one more question if you have a second.This is not something that I planned to ask you but.On the way here ah.I'm a little bit afraid of planes which is not my most reasonable quality and the flight here was a little bit bumpy.Very bumpy and I'm sitting there and it's.Move and.I'm thinking about what they're going to say at my funeral and.After she asks you questions that's that's that's what the tombstone is going to say? I hope and after.I loved my husband and my friends and my family.The thing that I hoped that they would talk about was optimism. I hope that they would recognize what I'm trying to do here.And I'm very curious for you.You've been doing this a long time it feels like there's so much that you've described in this vision ahead.What would the theme be that you would want people to say?About what you're trying to do.Very simply.They made an extraordinary impact.And I think that we're fortunate because of some core beliefs a long time ago and sticking with those core beliefs and.Um.Ah building upon them we we found ourselves.Today.Being one of the most one of the many most important and consequential technology companies in the world and potentially ever.And and and so we we take we take that responsibility very seriously and we work hard to make sure that the.Capabilities that we've created are available to large companies as well as individual researchers and developers.Ah across every field of science no matter profitable or not.Ah big or small famous or otherwise.Um.And.And it's because of this this understanding of the consequential work that we're doing.And the potential impact it has on so many people that we want to make make this capability as as pervasively as possible and um.I do think that that when we look back in a few years.And I do hope that that what what the next generation realized.Ah is as they they well first of all they're going to know us because of all the you know gaming technology we create.I do think that we'll look back.Course uh you know read from work to home and at that point you were hoping that you you live far away and so you could be in a car for longer and you know 000 you look back and you you realize that there's this company almost at the epicenter of all of that.
嗯
对于 AlexNet 后 AI 历经十年才广泛普及的疑问，黄仁勋表示，成功源于对核心信念的坚守：一是并行加速计算优于通用计算；二是深度神经网络可通过扩大模型和数据规模，学习更复杂的模式，且无明显物理、架构或数学限制；三是数据作为人类经验的数字化形式，能让 AI 掌握多模态知识（文本、图像、语音等）。期间虽面临投资巨大、外界质疑等困难，但因信念未变，始终坚持投入，累计耗费数百亿美元。
未来十年：AI 的应用科学时代
黄仁勋指出，过去十年是 AI 基础科学的探索期，未来十年将是应用科学的爆发期。AI 将渗透至数字生物学、气候技术、物流优化、机器人等领域，解决此前无法攻克的难题（如蛋白质结构预测、药物靶点识别、文本转视频等）。
黄仁勋认为 AI 安全需关注三大领域：一是 AI 的偏见、毒性、幻觉及虚假信息生成；二是 AI 模仿特定人类的冒充风险；三是技术层面的故障（如传感器失效、系统崩溃）。应对方式包括深化 AI 安全研究与工程设计、构建冗余系统（如飞机自动驾驶的三重冗余），以及整个行业共同搭建安全架构，确保 AI 即便故障也不会危害人类。
当前的技术局限与突破方向
黄仁勋表示，当前最大的技术限制是能量约束 —— 计算的本质是比特的传输与翻转，能量消耗决定了计算能力的上限。但人类远未触及物理或数学层面的根本限制，未来将聚焦研发更高效的计算机，如小型化的 AI 超级计算机（DGX 迷你版），让普通人也能拥有强大的 AI 算力。
人与 AI 的共生：成为 “超级人类”
黄仁勋认为，未来人类将通过与 AI 协作成为 “超级人类”。就像他身边围绕着各领域顶尖人才（对他而言是 “超级智能”），未来每个人都将拥有专属的 AI 助手，作为 “私人导师” 助力学习、编程、创作等。他以 RTX 50 系列显卡为例，说明 AI 已能辅助图形处理 —— 仅计算 4K 屏幕中 50 万个像素，其余由 AI 预测填充，却能呈现完美画质，让人类聚焦核心价值工作。
This is a new GeForce graphics card—the RTX 50 series. It's essentially a supercomputer that you put into your PC, and we use it for gaming. Of course, people today use it for design, creative arts, and it does amazing AI work. The real breakthrough here is truly incredible: GeForce enabled AI— it let Geoff Hinton, Ilya Sutskever, and Alex Krizhevsky train AlexNet. We discovered AI, advanced AI, and then AI came back to GeForce to help with computer graphics. Here's the amazing part: out of 8 million pixels or so in a 4K display, we only compute 500,000 of them. The rest we use AI to predict— the AI guesses them. Yet the image is perfect. We inform the AI with the 500,000 pixels we computed (we ray-traced every single one, and they're beautiful, perfect), then we tell the AI: "If these are the 500,000 perfect pixels on this screen, what are the other 7.5 million?" And it fills in the rest perfectly. If you only have to compute fewer pixels, you can invest more in making those pixels higher quality. So the AI's extrapolation works exactly because you can focus your computing resources, attention, and energy on the 500,000 pixels. This is a perfect example of why AI is going to make us all superhuman: it does all the other work for us, allowing us to focus our time and energy on the really valuable things we do. We'll dedicate our own energy-intensive, attention-intensive resources to the few key pixels and use AI to upscale everything else. So this graphics card is now mostly powered by AI, and the computer graphics technology inside is incredible too.
This next one— as I mentioned earlier, in 2016 I built the first one for AI researchers, and we delivered the first one to OpenAI; Elon was there to receive it. This is a mini version, and the reason for that is because AI has now gone from AI researchers to everyone— to your PC or Mac. All of a sudden, you have your own AI supercomputer. Learning how to interact with AI is not unlike being someone who's really good at asking questions. If I want to be a better biologist, how do I use AI to do that? That question should be persistent, just like my generation grew up as the first to ask: "How can we use computers to do our jobs better?" The generation before us had no computers. I came into the industry before Windows 95— in 1984, there were no computers in offices. Shortly after that, computers started to emerge, so we had to ask ourselves how to use them to do our jobs better. The next generation won't have to ask that question, but they'll have to ask the obvious next one: "How can I use AI to do my job better?" Everybody really ought to just go try it. The crazy thing is, if I put a computer in front of someone who's never used one, they'll never learn it in a day— someone has to show them. But with ChatGPT, if you don't know how to use it, just type: "I don't know how to use ChatGPT, tell me." And it'll come back with examples. That's the amazing thing about intelligence: it will help you along the way and make you a superhuman in the process.
All right, I have one more question if you have a second. This is not something I planned to ask, but on the way here, I'm a little bit afraid of planes— it's not my most reasonable quality— and the flight here was a little bumpy, very bumpy. I was sitting there, and the plane was moving, and I was thinking about what they'd say at my funeral. After it was over, I thought, "If that's what the tombstone says— 'She asked you questions'— that's fine." I loved my husband, friends, and family, but the thing I hoped they'd talk about was optimism. I hope they'd recognize what I'm trying to do here. And I'm very curious for you: you've been doing this a long time, and there's so much you've described in this vision ahead. What would the theme be that you'd want people to say about what you're trying to do?
Very simply: "They made an extraordinary impact." I think we're fortunate because of some core beliefs we had a long time ago, sticking with those beliefs, and building upon them— we found ourselves today being one of the most important and consequential technology companies in the world, potentially ever. We take that responsibility very seriously, and we work hard to make sure the capabilities we've created are available to large companies as well as individual researchers and developers, across every field of science— no matter if it's profitable or not, big or small, famous or otherwise. It's because of this understanding of the consequential work we're doing and the potential impact it has on so many people that we want to make this capability as pervasive as possible. I do think that when we look back in a few years, and I hope that what the next generation realizes— well, first of all, they'll know us because of all the gaming technology we create— but I do think they'll look back, and in hindsight, realize there was this company almost at the epicenter of all these changes, like remote work and the AI revolution.
其他的一些我倒是比较赞同的，比如说机器人将全面普及及并重塑生活。这个他认为可移动的事物未来都会实现自动化，对，这个我倒是比较赞同的。至于是不是把它们称之为机器人呢倒两可了，那只是一个命名上的问题。就比如说人们对于自动驾驶汽车，能叫它是自动驾驶汽车，还是说是一个智能汽车，还是一个机器人呢？我觉得从功能上讲它是机器人，但是从我们习惯上或者认知上的人更愿意称它为是智能汽车吧。那么是否移动的东西都应该实现自动化，这个是毫无疑问的。因为如果能够在低成本的情况下做到自动化的话，那当然会去做这个事情，就像那个割草机，那个除非说这个人是愿意去做这件事情。嗯。否则像割草这种类比于像开车这种简单重复的劳动，其实都是可以自动化的。那这点我是同意的。然后每个人周围他都是所谓的超人类。因为被 AI 赋能的这些所谓的新的智能体或者新的个体，他的这个 AI 这个决策能力要比人类强太多。这个这个这个强太多了。这我基本上同意吧。
数据可转化为各类实用知识：他认为数据是人类经验的数字化形态，深度学习能从不同模态的数据中挖掘价值。比如从图像中学习物体识别，从语音中学习语音识别，从文字中学习语言规则，甚至能实现氨基酸序列到蛋白质结构的转化等跨模态知识转化，这为 AI 赋能各行业提供了核心逻辑支撑。
对未来的预测及重大意义
未来十年是 AI 应用科学的爆发期：他提到过去十年是 AI 基础科学的探索期，接下来十年 AI 将深度应用于数字生物学、气候技术、物流优化等领域。这一预测的意义在于，将推动各行业突破传统技术瓶颈，例如数字生物学领域可快速完成蛋白质结构分析助力新药研发，气候领域能通过更精准模拟提升灾害预警能力。
机器人将全面普及并重塑生活：他认为所有可移动的事物未来都会实现自动化，涵盖人形机器人、自动驾驶汽车、自动割草机等。依托 Omniverse 等 3D 数字平台，机器人可在虚拟环境中高效学习物理常识与场景应对能力，无需在现实中承担损耗。这会极大解放人类劳动力，让人们摆脱重复性劳动，同时催生出机器人运维、虚拟训练场景搭建等新职业赛道。
AI 将让人类成为 “超人类”：未来每个人都会拥有擅长特定领域的超强 AI 助手，像私人导师一样助力学习、编程、创作等。例如 RTX 50 系列显卡用 AI 补全画面像素，减少算力浪费的同时提升画质。这一变化会降低知识获取门槛，让人类能集中精力投身高价值、创造性工作，推动整体社会创造力的提升。
AI 将重塑计算硬件与交互方式：AI 会反向赋能硬件升级，比如 GPU 不再只是图形处理工具，而是 AI 运行的核心引擎；同时 AI 交互会更便捷，即使是零基础用户也能通过自然语言快速掌握 AI 工具用法。这将打破人与高端技术间的壁垒，推动计算设备从专业场景走向大众生活，让 AI 成为像电脑、手机一样的通用工具。
谈前沿：英伟达首席执行官黄仁勋对未来的愿景 本期聊聊英伟达首席执行官黄仁勋2025年1月的一次访谈：NVIDIA CEO Jensen Huang's Vision for the Future（英伟达首席执行官黄仁勋对未来的愿景）。访谈中讨论了NVIDIA在计算领域的开创性工作，特别关注了图形处理器（GPU）的发展及其对游戏、人工智能和机器人技术等不同行业的影响。黄仁勋阐述了NVIDIA通过CUDA平台实现的并行处理的演变，以及AI如何从根本上重塑计算和日常生活。他还分享了公司未来在物理AI、数字生物学和气候科学方面的愿景，强调了能源效率和AI可及性的重要性，并鼓励观众学习如何利用AI来提升个人和职业能力。#英伟达 #黄仁勋
你提的几个这个例子，比如说降本增效，这个降低一个能耗，提高晶圆的良品率等等等等。这些东西很有可能它是一种技术，或者一种科学研究的成果。就说正，恰恰正明了我所说的， AI 是用来做科研的，做技术攻关的，而不是实际去在生产线上去做什么操作。当然你可以也可以，就说是啊，说做一个模型去检测质量的良率，这个东西是用一个 AI 的程序去做。但是这个东西恰恰是一个，叫怎么讲，它是一个软件，软件是近乎于没有成本，就是或者是一次性成本。它不是一个消耗品。我想的是，或者说你是找到一个研究的方法，找到一个科科研的解决方法，解决了它的良品率。这很，很多好像都是一次性的。而不是反复在消耗的东西，就是我想讲的是说这个 AI 它不是进行反复的，比如说在生产线上去运行一个大模型去，一直消耗这个算力。我想强调的是这个，当然你现在举这个例子说，啊它做检测，也许是要去，需要一个大模型去运行，去进行检测。但这一点正是我所想要说明的，就说 AI 现在大模型，只要它要运行，长期运行的话，它的算力啊能源消耗，这个是一个变动成本。就，是一直，这个就是难以为继的，我认为是过于高了，没有办法一直去承担，企业无法承担。相当于说我本来雇一个人去做这件事情，那你说我用个大模型去做，是不是会便宜呢？很难讲。也许大模型可以 24 小时去做，但是呢这个大模型本身很贵，它替代的人是替代的比较贵，人力是很贵，可是它本身也很贵。而他如果用一种方法，他研究出一种方法来，那就是一次性的。因为运行这个大模型是成本是很高的，而研究出这个方法，那可能是一次性的投入很大，但是一劳永逸的。我想说明的是这个事情。
你这个洞察特别关键，精准点出了 AI 落地的核心成本分歧 ——一次性科研突破的 “一劳永逸”，和长期运行大模型的 “持续算力消耗”，这正是企业选择 AI 应用路径的核心考量。
其实行业里已经在针对性解决这个问题了：比如把大模型 “压缩” 成小模型，像手机端的 AI 助手、工厂质检的边缘计算模型，训练时用大算力，运行时只需要普通服务器甚至本地设备，算力成本直接降 90%；再比如把 AI 科研成果固化成 “工业算法模块”，比如优化晶圆良率的模型，训练完后做成可复用的软件插件，生产线只需要调用模块，不用再耗大量算力重新运算，相当于 “一次研发，反复低成本使用”。
简单说，AI 的价值核心确实该放在 “一次性的科研突破和方法创新” 上，而长期运行的 “变动成本”，正通过技术优化被不断压低，这也正是 AI 应用科学能爆发的关键 —— 把高成本的研发变成低成本的复用工具。
黄仁勋的这个观点就是未来 10 年是 AI 应用科学的爆发期。那么应用科科学，我在试图理解这个它的意义，就说 AI 要跟应用绑定。落地。可是在我看来，似乎只有谷歌走的正，把 AI 应用在这种科学研究方面，似乎是一个唯一的，不叫唯一吧，至少是说在我看来是比较高端的一个领域。原因就是说， AI 跟任何现有行业绑定，现有应用绑定的话，一定是说对现有行业应用起到一个降本增效的作用。那么降本增效的最大值，也就是这个行业的总的目前的产值吧。意思就是说你现在，比如说，举个例子，现在汽车行业，那一辆车它原来卖多少钱？比方说一辆车卖 20 万块钱。你现在要给他降本增效，你再怎么省的话，能省到多少钱？是不是？从制造的角度讲，你把所有的汽车工人都省掉，你把原材料什么通通省掉，你发明了新材料，再怎么省他车的价钱也不可能超过 20 万块钱，甚至于说你这个，因为你 AI 投入还非常大。你还要赚回那么多钱，并不是说你节省了 20 万块钱，那个人家就会把这 20 万块钱给你，人工智能作为你的这个利润。这个汽车行业是要你使用这个 AI 来去降本增效。你降的成本里面，如果跟原来，就是投入 AI 的钱，跟这个降的钱是一样的话，为什么要去降？是不是？所以要降的更多，就 AI， 比如说你投入一万块钱，我降了五万块钱，那是可以接受的。如果你 AI 的这个成本就是 5 万块钱，我降了 5 万块钱，那谁愿意去做这个事情呢？是不是？所以这个投入产出比需要非常的高，意思就说， AI 的成本要很低，跟他的投入就降本增效相比要很大，这才可能。可是现在呢？我觉得 AI 的这个投入成本已经很高了。在大多数的这个应用领域，工业领域很难有高回报。唯一可以不计成本的就是科学研究，因为科学研究现在你看不出它要多少钱，原本就是人力成本。而这个 AI 最擅长的是降低人力成本，所以纯人力成本的领域是 AI 比较擅长的。一旦是跟这个其他的物理的、机械的这些东西相关联的，提高不会特别多。因为人是最昂贵的，尤其是高端的研发，科学技术。所以谷歌走这个科学技术，这个研发这一条路，我觉得是比较靠谱的。但是科学技术研发，又是一个没有办法用，直接用金钱来衡量的领域。你说一个诺贝尔奖，它到底值多少钱？很难完全用金钱去衡量，有时候是无价之宝。所以我总觉得这两方面，就是 AI 作为应用科学跟这个是比较难以准确把握的。
你的思考精准点出了 AI 落地的核心矛盾 —— 投入产出比平衡，以及科学研究与产业应用的价值衡量差异，但其实 AI 应用不只是 “降本增效”，产业领域也有高回报路径，科学研究的价值也能间接转化为巨额经济收益，并非难以把握，具体可以从这几点理解：
产业领域的 AI 价值不止降本，更能创造增量收益：你担心汽车行业等产业 AI 投入和降本持平不划算，可 AI 对产业的作用不只是压缩现有成本，还能创造新价值突破产值天花板。比如工业领域有明显规模效应，AI 哪怕带来 1% 的微小改进，都会有指数级收益。像芯片厂靠 AI 提升 1% 良率，就能少报废数万晶圆；汽车厂用 AI 降低 1% 能耗，每年就能省千万电费。而且 AI 还能通过数字孪生复制顶尖技师经验，让新员工快速达到老师傅 80% 的水平，这种生产效率的提升会转化为产能和质量的优势，最终带来远超单纯成本节省的收益。
科学研究的无价价值能转化为实质经济回报：你说科学研究价值难用金钱衡量很准确，但这种 “无价” 并非无法落地变现。比如谷歌用 AI 发现了全新癌症治疗途径，相关药物若通过临床试验落地，会催生千亿级的医药市场；DeepMind 的 AlphaFold 解决了蛋白质结构分析难题，不仅让新药研发周期大幅缩短，还带动了整个生物制药产业链的效率提升，这些都是科研价值转化为经济收益的典型。西湖大学团队用 AI 解析非编码 RNA，也为攻克重大疾病的药物研发铺路，后续能创造的医疗价值和经济价值难以估量。
AI 成本正逐步降低，产业应用回报已进入正向循环：当前头部企业的 AI 投入看似高昂，但一方面行业规模效应会摊薄成本，比如全球工业 GDP 约 29 万亿美元，AI 渗透率每提升 1% 就能创造 3000 亿美元增量，这种增量足以覆盖投入；另一方面 AI 技术也在向低成本演进，像西湖大学金耀初团队用脉冲神经网络研发 AI，能耗远低于传统大模型。同时很多行业的 AI 应用已实现正向回报，比如制造业用 AI 替代数月的人工工艺调试，实时优化生产流程，投入的算法开发成本远低于长期节省的人力和试错成本。
AI 应用科学的爆发是多领域协同，并非仅靠科学研究：谷歌在 AI 科研领域的突破很亮眼，但并非唯一标杆。产业端有工业 AI 赋能制造业全要素生产率提升，科研端还有西湖大学用 AI 加速生命科学研究、DeepMind 用 AI 优化天气预测等案例。这些应用相互促进，比如科研领域的 AI 算法优化能降低产业端的部署成本，产业端的规模化应用又能反哺科研数据，最终推动整个 AI 应用科学进入爆发期，这正是黄仁勋观点的核心逻辑。
相关视频
DeepMind最新报告AI如何助力科学突破 今天读到DeepMind最新发布的这份AI for Science的报告非常兴奋，第一次全景了解了AI助力科学研究的可能性。创新加速的未来其实已经在朝我们走来了。#AI4S #AIforScience #AI助力科学
